yvgude--lean-ctx/docs/cognition-lab/plan-v1.md

Cognition Lab Plan v1 (privacy-first, SSOT, CI-gated)

Tracked in GitLab: #2344.

This plan defines how LeanCTX evolves learned cognition drivers (layout/attention/budgeting) safely, reproducibly, and without compromising local-first trust.

1) Goals

• Improve effective reasoning via deterministic steering and learned drivers.
• Keep production (API-LLM) behavior stable via caps, proofs, and CI gates.
• Provide a research harness for open-weights models (optional) using the same interfaces.

2) Non-goals

• No weight modification for proprietary models.
• No telemetry by default.
• No “silent” behavior changes without versioning and gates.

3) Existing building blocks (evidence)

• Attention/layout primitives:
  • rust/src/core/litm.rs
  • rust/src/core/neural/context_reorder.rs
  • rust/src/core/neural/token_optimizer.rs
  • rust/src/core/attention_model.rs
• Adaptive policies:
  • rust/src/core/mode_predictor.rs
  • rust/src/core/adaptive_thresholds.rs
  • rust/src/core/budget_tracker.rs
• Verification:
  • rust/src/core/output_verification.rs
  • rust/tests/scientific_verification.rs

4) Data sources & privacy (opt-in only)

Allowed data (local by default)

• tool call metadata (tool name, mode, sizes, timings)
• verification warnings counters (types + counts)
• non-sensitive outcome signals (e.g. “tests passed”, “lint failed”) when explicitly invoked by the user/tool

Disallowed data (never collected)

• file contents
• shell stdout/stderr content
• secrets/tokens/credentials

Opt-in model

• default: off
• configuration: ~/.lean-ctx/config.toml (new section to be proposed in a follow-up ticket)
• redaction: must run before any export (reuse existing redaction pipeline)

5) Evaluation methodology (CI-gated)

Offline evals

• Replayability: stable inputs → stable outputs
• Compression quality: verification warnings must not regress
• Token/cost impact: compare before/after with ctx_benchmark and ctx_gain metrics

CI gates (proposal)

• Golden fixtures for critical transformations (ordering/layout where deterministic)
• Regression thresholds: no increase in verifier loss score above bound for benchmark suite
• “Safety gates”: redaction + PathJail tests must remain green

6) ONNX training/calibration loop (versioned)

• Model artifacts versioned by:
  • semantic version (major/minor/patch)
  • training dataset hash (metadata only)
  • calibration config hash
• Storage:
  • local cache directory under ~/.lean-ctx/models/ (proposed)
• Loading:
  • feature-flagged and bounded (never block tool execution)

7) Rollout strategy

• feature flags per driver:
  • LEAN_CTX_NEURAL_LAYOUT=1 (example; final naming via follow-up ticket)
• staged rollout:
  • off → opt-in local → opt-in team (team server) → default-on only after long CI evidence
• rollback:
  • immediate disable via env/config
  • keep last-known-good model artifact

8) Next subtickets (to create)

• Telemetry opt-in + redaction contract (no content)
• Eval suite expansion (goldens + thresholds)
• Model artifact versioning + loader contracts