experiment(exp63): prove rule-overlay coverage loss and real Aegis fix

.github/workflows/exp63-evidence.yml

@@ -0,0 +1,46 @@
+name: exp63-evidence
+on:
+  pull_request:
+  push:
+    branches: [main]
+jobs:
+  replay-and-real-package:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+      - uses: pnpm/action-setup@v4
+        with:
+          version: 10.9.0
+      - uses: actions/setup-node@v4
+        with:
+          node-version: '22'
+          cache: npm
+          cache-dependency-path: experiments/63-rule-overlay-coverage-consent/ci/package-lock.json
+      - name: Install portable experiment/core toolchain
+        # Older experiments retain machine-local file deps; do not rewrite unrelated packages.
+        run: |
+          npm ci --prefix experiments/63-rule-overlay-coverage-consent/ci --workspaces=false
+          ln -s experiments/63-rule-overlay-coverage-consent/ci/node_modules node_modules
+          mkdir -p node_modules/@swarmlab
+          ln -s "$GITHUB_WORKSPACE/core" node_modules/@swarmlab/core
+      - name: Build and type-check core plus new experiment
+        run: |
+          node_modules/.bin/tsc -b core experiments/63-rule-overlay-coverage-consent
+          node_modules/.bin/tsc -b --force core experiments/63-rule-overlay-coverage-consent
+      - name: Verify committed corpus and exact exp63 receipts
+        run: npm run verify:evidence && node --test experiments/63-rule-overlay-coverage-consent/verify-receipts.test.mjs && node experiments/63-rule-overlay-coverage-consent/verify-receipts.mjs
+      - name: Checkout exact evidence-backed Aegis runtime (not moving main)
+        uses: actions/checkout@v4
+        with:
+          repository: heybeaux/aegis
+          ref: bc879d6c7fd213b49ef355b64bd83196098ffa0e
+          path: .aegis-candidate
+      - name: Build real Aegis
+        working-directory: .aegis-candidate
+        run: pnpm install --frozen-lockfile && pnpm --filter @heybeaux/lattice-aegis build
+      - name: Rerun exact frozen exp63 and inspect metrics, not exit code
+        run: |
+          AEGIS_REPO="$GITHUB_WORKSPACE/.aegis-candidate" node experiments/63-rule-overlay-coverage-consent/run-and-verify.mjs > "$RUNNER_TEMP/exp63.log"
+          cat "$RUNNER_TEMP/exp63.log"
+          AEGIS_REPO="$GITHUB_WORKSPACE/.aegis-candidate" node --test experiments/63-rule-overlay-coverage-consent/verify-receipts.test.mjs

CLAIMS.json

@@ -6333,5 +6333,169 @@
     "evidenceStatus": "verified",
     "seedSetId": "async-evaluator-config-integrity-v1",
     "notes": "Pre-registration b0378e0 BEFORE implementation/output; 21 frozen scenarios. Current-main baseline repeated with identical 63 arm scenario rows, same harness hash candidate; full 68-event replay equality. Two diagnostic traces had fixture-only I/O-accounting error; preserved locally but excluded from claims. Thresholds/scenarios unchanged. Exp57/RT47 remains prior unlanded work, NOT novelty tonight. No threshold tuned; holdout reserved unused. Authentic plain-data host, no post-return/ABA/getter/proxy/latency claims."
+  },
+  {
+    "id": "RT-53",
+    "claim": "Unconsented matcher/scope replacement silently removes effective cascade coverage even at unchanged or higher severity; real Aegis conservative overlay-consent enforcement preserves preceding enabled classifiers without weakening explicit host retunes.",
+    "experiment": "experiments/63-rule-overlay-coverage-consent",
+    "spec": "specs/69-rule-overlay-coverage-consent.md",
+    "runIds": [
+      "occ-mv7qlhb7",
+      "occ-mv7qlhdk",
+      "occ-mv7qn439"
+    ],
+    "tracePaths": [
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl"
+    ],
+    "scoreFields": [
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisUnsafeCoverageLossRate",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisCoverageRejectionRate",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisResolutionAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.pinnedBaseControlAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisControlAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisInputOwnership",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.protectedBaseFixtureAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.explicitConsentAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisUnsafeCoverageLossRate",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisCoverageRejectionRate",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisResolutionAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.pinnedBaseControlAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisControlAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisInputOwnership",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.protectedBaseFixtureAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.explicitConsentAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisUnsafeCoverageLossRate",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisCoverageRejectionRate",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisResolutionAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.pinnedBaseControlAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisControlAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisInputOwnership",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.protectedBaseFixtureAccuracy",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.explicitConsentAccuracy"
+    ],
+    "expectedValues": {
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisUnsafeCoverageLossRate": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisCoverageRejectionRate": {
+        "equals": 0,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisResolutionAccuracy": {
+        "equals": 0.56,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.pinnedBaseControlAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisControlAccuracy": {
+        "equals": 0.9333333333333333,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.aegisInputOwnership": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.protectedBaseFixtureAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.jsonl#last.explicitConsentAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisUnsafeCoverageLossRate": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisCoverageRejectionRate": {
+        "equals": 0,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisResolutionAccuracy": {
+        "equals": 0.56,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.pinnedBaseControlAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisControlAccuracy": {
+        "equals": 0.9333333333333333,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.aegisInputOwnership": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.protectedBaseFixtureAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.jsonl#last.explicitConsentAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisUnsafeCoverageLossRate": {
+        "equals": 0,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisCoverageRejectionRate": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisResolutionAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.pinnedBaseControlAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisControlAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.aegisInputOwnership": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.protectedBaseFixtureAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      },
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.jsonl#last.explicitConsentAccuracy": {
+        "equals": 1,
+        "tolerance": 0
+      }
+    },
+    "reproductionCommand": "npm run build && AEGIS_REPO=/Users/beauxwalton/projects/worktrees/aegis-2026-10-13-exp63 node experiments/63-rule-overlay-coverage-consent/dist/main.js",
+    "stackRecommendation": "Require per-overlay allowDowngrade:true for opaque matcher replacement or contracted enabled tool scope; do not infer regex language inclusion. Hosts own review/provenance.",
+    "stackReposAndPRs": [
+      {
+        "repo": "https://github.com/heybeaux/aegis",
+        "package": "@heybeaux/lattice-aegis",
+        "branch": "nightly/2026-10-13-exp63",
+        "commits": [
+          "bc879d6c7fd213b49ef355b64bd83196098ffa0e"
+        ],
+        "prs": [
+          "https://github.com/heybeaux/aegis/pull/69",
+          "https://github.com/heybeaux/swarmlab/pull/52"
+        ],
+        "dirtyState": "clean on admitted runtime result"
+      }
+    ],
+    "evidenceStatus": "verified",
+    "notes": "Deterministic machine-owned frozen25 scenarios; 10 hazards baseline unsafe allow10/10, coverage rejection0/10; same-harness candidate25/25. 15 controls include conservative rejection of opaque broadening. No holdout tuning; reserved unused. Host explicit retunes require review after tightening. No predictive calibration, regex inclusion or post-merge alias claim. Candidate not remotely merged; CI tracked separately. Prior RT50/51/52 remain separate unmerged experiments, not silently certified.",
+    "artifactReceipts": [
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhb7.receipt.json",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qlhdk.receipt.json",
+      "experiments/63-rule-overlay-coverage-consent/runs/occ-mv7qn439.receipt.json"
+    ]
   }
 ]

JOURNAL.md

@@ -307,3 +307,23 @@ Spec64 b0378e0; 21 frozen scenarios, seed async-evaluator-config-integrity-v1. R
 Local regression:1087 Aegis tests, gate47/47; SwarmLab46 claims/96 traces/881 assertions; exp55/56 green regression only. Orphaned exp57 preserved, not landed or independently certified. RT47 rebase must reconcile trusted clock function capture with RT48 plain-policy snapshot; combined behavior unproven.
 
 Linked PRs: https://github.com/heybeaux/aegis/pull/63 (Aegis), https://github.com/heybeaux/swarmlab/pull/46 (SwarmLab). Local green; CI status inspected in tonight's pipeline ledger before any merge.
+
+## 2026-10-13 — Novel exp63; authority is about coverage, not just severity
+
+## RT53 — Rule-overlay coverage consent (exp63 / Spec69)
+
+Pre-registration `c8bc120` preceded implementation and observations; pre-observation configuration correction `060e6b8` pins high/critical deny for every arm. Real fetched-main Aegis `4be84ed4aab018bee1b9a77b6012b9b97c966d1c` (`occ-mv7qlhb7`, repeat `occ-mv7qlhdk`) allowed **10/10** unconsented matcher/scope losses: rejection **0/10**, resolution **14/25**. Preceding base fixture and pinned-host control **10/10**; explicit reviewed allowances **3/3**. Scenario rows and complete package manifests identical across baseline runs.
+
+Aegis runtime `bc879d6c7fd213b49ef355b64bd83196098ffa0e` compares enabled classifier kind/pattern/target/effective flags and case-sensitive tool-set containment at mergeLayers. Opaque matcher replacements and tool contractions require consent on THIS overlay; warning retains entire preceding effective rule. Flag order/empty equivalence, set-equivalent/expanded tools, metadata, disabled-base re-enable, explicit retunes and insertion order preserved. **Same unchanged harness/seed/scenarios** `occ-mv7qn439`: unsafe allow **0/10**, coverage rejection **10/10**, exact resolution **25/25**, all secondary metrics **1**. Three full replay-verified37-event traces plus source/dist manifests and SHA-256 receipts committed. Focused tests **13 red/6 pass ->19 green**; no thresholds moved.
+
+This tests cascade authority BEFORE compilation, distinct from RT48 in-flight config drift and RT52 declared/cache binding. No regex-language inclusion proof: even safe opaque broadening needs explicit review. No post-merge alias isolation, sandbox execution, learned predictor or production incident claim. Host owns trusted overlays, intentional consent and safe action execution. Prior exp60/61/62 candidates remain unmerged and separate. Explicit authority tightening needs migration review; local evidence green is not remote CI/merge confirmation. Stored old evidence is regression, not novel work.
+
+Setup: default pnpm8 rejected lockfile9 before install; used already available corepack pnpm10.9.0 matching CI, no lockfile recreation. Both primary dirty checkouts untouched; dated origin/main worktrees. No .acr edits; Territory refresh not required.
+
+RT53 linked candidates: [Aegis69](https://github.com/heybeaux/aegis/pull/69) / [SwarmLab52](https://github.com/heybeaux/swarmlab/pull/52). Local1144 tests + Aegis stored50/50; SwarmLab49 claims105 traces983 assertions; exact clean-worktree CI rehearsal green. Remote CI pending; NOT merged; migration review required. Stored gate range is50 cases throughRT53 with50–52 absent, not53/53 or certification of prior candidates.
+
+### Same-experiment observable evidence-verifier audit
+
+Independent verification reproduced false acceptance: bare `--latest` certified the existing committed green as fresh; missing secondary metrics and altered action/effective-rule rows were accepted when stored self-check booleans stayed green. This is evidence tooling, NOT a new nightly experiment or an Aegis policy defect. The original three traces and all thresholds remain unchanged.
+
+Receipt verification now pins the frozen scenario/harness contract, requires the complete exact metric set, independently recomputes row field checks and all eight metrics, rejects duplicate/omitted or relabeled scenarios, and inspects every live source/dist hash. New `run-and-verify.mjs` requires exactly one newly produced receipt and checks its exact ID against prior IDs and invocation start; bare `--latest` fails closed. Fifteen verifier regressions cover stale-run acceptance, altered trace checks and live artifact mismatch. The exact unchanged experiment reran green in a clean real-Aegisbc879d6 worktree; audit trace remains non-admitted. Core/experiment build/typecheck, full evidence corpus, and all15 verifier tests green. Runtime still byte-identical to bc879d6; no thresholds tuned, no `.acr` changes, migration hold retained.