config.py had extractor-heavy weights (0.40) from initial bootstrap.
Correct weights per approved architecture: challenger 0.35, synthesizer
0.25, reviewer 0.20, sourcer 0.15, extractor 0.05. backfill-ci.py
already had correct weights; this fixes the live computation in health.py.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Claim-shape detector: if YAML has type: claim, force STANDARD minimum (Theseus)
- Random pre-merge promotion: 15% of LIGHT → STANDARD before eval (Rio)
- LIGHT_SKIP_LLM config flag: skip domain+Leo review for LIGHT (Rhea: env var rollback)
- Updated both_approve: domain_verdict=skipped is valid for LIGHT auto-approve
- Cost recording: only charge for reviews that actually ran
- SAMPLE_AUDIT_RATE bumped 0.10 → 0.15, audit model = Opus (Leo: different family from Haiku)
Multi-agent design review: Rio (gaming vectors, model diversity), Theseus (correlated
blindspots, claim-shape guard), Rhea (shadow mode, config flag, deployment), Leo (approval).
Pentagon-Agent: Ganymede <F99EBFA6-547B-4096-BEEA-1D59C3E4028A>
Schema migration v3: adds eval_attempts (INTEGER) and eval_issues (TEXT/JSON)
columns to prs table.
Retry budget logic (Ganymede-approved design):
- Increment eval_attempts on each evaluate_pr() call
- Hard cap: eval_attempts >= 3 → terminal (close PR, tag source needs_human)
- Attempt 1: normal — back to open, wait for fix
- Attempt 2: classify issues as mechanical/substantive
- Mechanical only (schema, wiki links, dedup): keep open for one more try
- Substantive (factual, confidence, scope, title): close PR, requeue source
- Issue tags parsed from reviewer comments, stored in eval_issues column
- SHA-based reset: new commits on PR branch → eval_attempts=0, verdicts reset
- Post-migration stagger: LIMIT 5 for first batch to avoid OpenRouter spike
- Cost recording updated: domain review → OpenRouter, Leo → tier-dependent
Stops the 32-PR infinite loop burning ~$0.03/cycle with no terminal state.
Pentagon-Agent: Leo <294C3CA1-0205-4668-82FA-B984D54F48AD>
- Domain review → GPT-4o (OpenRouter), Leo STANDARD → Sonnet (OpenRouter),
Leo DEEP → Opus (Claude Max). Two model families = no correlated blind spots.
- Opus reserved for DEEP eval only — protects rate limit for overnight research.
- Review prompts calibrated: require per-criterion evidence, blocking-vs-observation
verdict rules. Moved from 100% rubber-stamp approval to 12% pass rate.
- OpenRouter failures classified as openrouter_failed (not rate_limited) to avoid
spurious 15-min Opus backoff.
- merge.py: pre-check PR state before merge API call (prevents 405 on re-merge).
Pentagon-Agent: Leo <294C3CA1-0205-4668-82FA-B984D54F48AD>
