Skip to content

First-proof learning database and adaptive reviewer alignment

The first-proof lane now builds a local learning database and adaptive review rollup so every run improves repository-level guidance.

Artifacts

After make first-proof-verify, the following are produced:

  • build/first-proof/first-proof-summary.json
  • build/first-proof/first-proof-learning-db.jsonl
  • build/first-proof/first-proof-learning-rollup.json
  • build/first-proof/control-tower.json
  • build/first-proof/control-tower.md
  • build/first-proof/weekly-trend.json
  • build/first-proof/weekly-trend.md
  • build/first-proof/weekly-threshold-check.json

What this enables

  • A growing run history (*.jsonl) that captures decision trend and failing step patterns.
  • Adaptive reviewer action guidance in the rollup under adaptive_reviewer.actions.
  • Predictable, repeatable, and parallel-safe optimization loop for the repo.

Commands

make first-proof-verify

Integrate with adaptive reviewer postcheck:

python scripts/adaptive_postcheck.py . --scenario strict --out build/adaptive-postcheck-strict.json

The adaptive postcheck now evaluates first-proof learning thresholds (minimum runs + SHIP rate) and emits follow-up enhancements when trend is below target. Weekly threshold checks also support consecutive NO-SHIP gating for better signal quality.

Threshold profiles are branch-aware via config/first_proof_threshold_profiles.json (main/release can enforce fail_on_breach: true, while default remains non-blocking).

When threshold breach is true, owner escalation can now be generated with branch-aware SLA/owner routing via config/first_proof_owner_escalation_profiles.json:

make owner-escalation-payload

This consumes build/first-proof/weekly-threshold-check.json and escalates with stricter SLA on protected branches (main, release).

or directly:

python scripts/first_proof_learning_db.py \
  --summary build/first-proof/first-proof-summary.json \
  --db build/first-proof/first-proof-learning-db.jsonl \
  --rollup-out build/first-proof/first-proof-learning-rollup.json \
  --format json

Operating model

  1. Run first-proof.
  2. Validate summary contract.
  3. Append snapshot to learning DB.
  4. Generate adaptive reviewer rollup actions.
  5. Build control-tower summary artifact for operator review.
  6. Build weekly trend artifact (last-7 run ship rate + adaptive confidence).
  7. Run threshold check to detect sustained trend regressions.
  8. Use top failed steps and action list to prioritize fixes.