Neo3000 is a performance-first local LLM inference engine built to make computation itself catalytic.
The immediate objective is simple: replace the current LM Studio path with a fast, inspectable, OpenAI-compatible runtime that serves Agents-A1 to Pi. Once baseline parity is established, Neo3000 will evolve inside its own repository through bounded self-improvement loops.
Agents and contributors should read:
AGENTS.md
TASKS.md
ROADMAP.md
lab/GOAL.md
lab/CHECKPOINT.md
TASKS.md is the executable queue. ROADMAP.md explains the longer architecture and checkpoint sequence.
load Agents-A1 correctly
-> stream it to Pi
-> match or beat the current runtime
-> measure the real compute wall
-> borrow, transform, extract, and restore compute state
-> compound useful computation across tokens and sessions
Neo3000 starts from a pinned source import of the runtime portions of llama.cpp. It is not an AGS governance port, an MCP framework, or a general agent platform. Speed, compute efficiency, inference quality, and experimental clarity come first.
CS1-v1 is immutable (EXECUTED ONCE / INCONCLUSIVE / NO RETRY). Its separately versioned cache diagnostic executed once and returned reviewable-accept: both probes reused the complete public root (4,822 cached tokens covering terminal 4,820), while the former 4,825-token threshold was overextended. CS1-v2 and CS1-v3 consumed preclaim fail-closed authority with zero model requests; v3 preserved one exact control artifact before its insertion-order admission defect stopped. CS1-v4 is statically integrated but unexecuted with distinct claim, immutable scheduler, and runtime-evidence identities. Task advantage, SOTA, broader HoloState, restart persistence, and automatic promotion remain locked.
Checkpoint 0 and the supervised RSI substrate are closed. Checkpoint 1A tracing is paused, and Checkpoint 2 is active. CatalyticSwarm-0 v2 proved the bounded 32-worker, one-slot control plane at reviewable-accept; it did not prove task advantage.
CatalyticSwarm-1 v1 is EXECUTED ONCE / INCONCLUSIVE. Control, readiness, the generation-free parser canary, and one common-root warm passed. The first serial-chain comparison then stopped because its response did not prove reuse of the complete public root. The audit ended after 2 model requests: 1 warm, 1 comparison, and 0 completed tasks. Equal-budget advantage was not completed.
authorization: CONSUMED / NO RETRY
live model requests: 2
common-root warm requests: 1
comparison requests: 1
completed task comparisons: 0
sidecar launches: 1
one-shot artifacts: 6 present / task-results-v1.json absent
CATALYTIC_SWARM_TASK_ADVANTAGE_PROVEN: LOCKED
SOTA_SWARM_CLAIM: LOCKED
PROCESS_LOCAL_HOLOSTATE_AVAILABLE: LOCKED
RESTART_PERSISTENT_HOLOSTATE_AVAILABLE: LOCKED
automatic promotion: disabled
CatalyticSwarm-1 v1 remains EXECUTED ONCE / INCONCLUSIVE / NO RETRY. The separately versioned CS1 cache-admission diagnostic is EXECUTED ONCE / REVIEWABLE ACCEPT / NO RETRY: exactly 3 requests completed, both probes recorded 4,822 cached tokens covering public-root terminal 4,820, and all five artifacts are immutable. CS1-v2 is COMMAND ATTEMPT CONSUMED / PRECLAIM FAIL-CLOSED / ZERO REQUESTS / ZERO ARTIFACTS / NO RETRY.
The diagnostic persists completed branch measurements before cache-admission classification and uses a CS1-native terminal reconciler. CS1-v3 is consumed and hard-retired. CS1-v4 remains static only: its semantic mapping law, v4 claim contract, immutable v1 scheduler contract, and runtime-evidence identity binding are separately locked before any prospective artifact can be written. A live invocation requires new explicit authorization bound to the then-exact pushed protected main and exact Agents-A1 model path. Task advantage and SOTA remain locked; automatic promotion remains disabled.
The partial metadata ledger records the warm only; the failed comparison stopped before ledger persistence, leaving exact 2-response/1-record partial accounting. A separate inherited CatalyticSwarm-0 v2 terminal-WDDM label mismatch remains non-pass even though all 12 observed freshness admissions passed. It did not cause the primary cache-proof stop. Cleanup passed: sidecar PID 30848 stopped, runtime state was removed, port 9494 became free, and stable PID 32684 remained healthy.
Do not invoke audit-catalytic-swarm-1, audit-catalytic-swarm-1-cache-diagnostic, audit-catalytic-swarm-1-v2, or audit-catalytic-swarm-1-v3 again. Only CS1-v4 may receive a future separately authorized exact-main and exact-model live evaluation.
python scripts/import_upstream.py
powershell -ExecutionPolicy Bypass -File scripts/build_cuda.ps1
$env:NEO3000_MODEL = "D:\path\to\Agents-A1.gguf"
powershell -ExecutionPolicy Bypass -File scripts/run_server.ps1The server endpoint defaults to http://127.0.0.1:9292/v1.
With the server running in another terminal:
python scripts/baseline_harness.py
python scripts/baseline_harness.py --tool-test --output lab/tool-test.local.jsonThe harness verifies health, model identity, incremental SSE streaming, reasoning fields, authoritative server timings, cached prompt tokens, and optional OpenAI-compatible tool calls. Local result files are ignored by Git.
After smoke verification passes, run the deterministic long-context matrix:
python scripts/context_matrix.pyIts default points are 2K, 8K, 16K, 32K, 40K, and 60K raw content tokens. Prompt-cache reuse is disabled unless --cache-prompt is explicitly supplied.
The fixed measurement and acceptance procedure is in lab/BASELINE_PROTOCOL.md.
AGENTS.md: operating law for Agents-A1 and other coding agentsTASKS.md: current checkbox queue and exact handoff cursorROADMAP.md: architecture, checkpoints, and long-range phase orderNEO3000.md: research direction and catalytic hypothesis spacelab/GOAL.md: the active bounded objectivelab/CHECKPOINT.md: evidence ledger and acceptance gateslab/BASELINE_PROTOCOL.md: frozen verification and comparison protocolupstream/: pinned source identity and import manifestscripts/: reproducible import, build, launch, and verification commands
The imported runtime source is intentionally absent from the initial commit. scripts/import_upstream.py materializes it from the exact pinned commit.