review-log

name	review-log
description	Read an entire runs/<id>/log.md and assess research narrative coherence, identify stuck hypotheses, detect overfitting signs, and write a review markdown. Use after a full run completes or when a loop has been stuck for many rounds.

/review-log

When to use

• A run of 10+ rounds has completed.
• User asks for a "research review" or "is the agent stuck?"

Inputs

• run_dir: path to runs/<id>/

Behavior

1. Run python -m cli.autoqec review-log <run_dir> to get structured stats (round count, Pareto size, top hypotheses, killed_by_safety count).
2. Read log.md fully.
3. LLM-reason:
   • Narrative coherence: does each round build on previous findings?
   • Stuck patterns: ≥3 rounds with near-identical hypotheses?
   • Overfitting signs: Δ_LER monotonically improving on train seeds but not holdout?
   • Safety-kill clustering: are VRAM or wall-clock kills bunched around one hypothesis family?
4. Write runs/<id>/review.md with: summary (5 sentences), top 3 concerns, recommended next actions.

Output

• Path to review.md.