@moolam/learning

Learning substrate contracts: turn-trajectory schema, outcome signals, baseline registry, and frozen eval slices for Track C.

Anti-cheat charter

Training environments must use the same production harness code path as live turns. The binding governance document is:

training/gym/charter.md

Executable replay-parity coverage: pnpm --filter @moolam/training-gym parity:check (and test).

Rewards computed under a diverged gym path are void and must not be promoted.

Trajectory critics (C3)

Pack-pluggable TrajectoryCritic / CriticScore contracts live in this package and are re-exported from training/critics/.

• Pure score(record) — no network / LLM on the default path
• Versioned rubricId + rubricVersion on every score
• CriticRegistry stores critics by id@version; pack oracles load from JSON manifests under training/critics/fixtures/pack-oracles/ (no domains/ imports)

pnpm --filter @moolam/learning test
# critics_interface + critics_pack_oracles cover determinism, hack ≤0, two CI oracles

Learning constitution

Binding governance law: docs/learning/CONSTITUTION.md.

One-surgery-per-stage promotion-candidate lint: pnpm --filter @moolam/learning surgery:check (green + seeded multi-surgery violation fixtures under training/eval/fixtures/promotion-candidates/).

Kill-switch operator runbook (constitution L4): docs/learning/KILL_SWITCH_RUNBOOK.md — copy-paste revert checklist, monthly safety-alignment drill schedule, verify once with parity:check + golden:replay.

Research intake (C7 breakthrough RFC)

New learning techniques enter through an eval-gated RFC — never ad-hoc trainer forks. Template + review workflow:

• docs/learning/research-intake/RFC_TEMPLATE.md
• docs/learning/research-intake/REVIEW_WORKFLOW.md
• docs/learning/research-intake/ADOPTION_CHECKLIST.md — approved → manifest → regenerate → micro-run → PROGRESS
• Worked example (GRPO G=8→G=6): docs/learning/research-intake/rfcs/RFC-2026-004-grpo-g8-to-g6.md
• Generator hook: docs/stages/tracks/_generator/track-c/research-intake-rfc.md

pnpm --filter @moolam/learning build
node --test packages/learning/tests/research_intake_rfc.test.mjs
node --test packages/learning/tests/research_intake_worked_example.test.mjs
node --test packages/learning/tests/research_intake_adoption.test.mjs
pnpm --filter @moolam/learning research-rfc-adoption:check

LLM-judge policy (C3)

Narrow lane for non-verifiable aspects only (tone, clarity). Binding governance: docs/learning/LLM_JUDGE_POLICY.md.

Published mirror: training/critics/llm_judge_lane.ts.

pnpm --filter @moolam/learning llm-judge-policy:check
pnpm --filter @moolam/learning llm-judge-gate:check

LLM judges must never score mastery math, citations, schema validity, or contract obligations — those stay with rule critics / pack oracles.

Held-out tone/clarity agreement fixtures live under training/eval/llm_judge_sets/ and are independent of critic human-label calibration.

Scripts

pnpm --filter @moolam/learning test
pnpm --filter @moolam/learning schemas:check
pnpm --filter @moolam/learning baselines:check
pnpm --filter @moolam/learning slices:check
pnpm --filter @moolam/learning surgery:check
pnpm --filter @moolam/learning hack:check
pnpm --filter @moolam/learning llm-judge-policy:check
pnpm --filter @moolam/learning llm-judge-gate:check
pnpm --filter @moolam/learning calibration:check