Improve RLCR efficiency with review gates, retry circuit breakers, and evidence ledgers #235

Summary

A completed RLCR session showed strong mainline control, evidence discipline, honest handling of negative results, and effective independent review. It also revealed several general opportunities to reduce review overhead and unstructured retries.

Suggested improvements

1. Add a pre-review completeness gate that rejects placeholder summaries unless they contain a verifiable blocker and attempted evidence.
2. Classify reviewer feedback as acceptance invariants, recommendations, or examples. Only correctness, safety, reproducibility, and acceptance requirements should be mandatory.
3. Add a retry budget and diagnostic circuit breaker. After two similar failures, require failure classification, environment checks, dependency checks, and single-variable retries.
4. Make least-privilege diagnostics the default through field allowlists, redaction, and pre-persistence scans.
5. Separate incremental narrative from a machine-readable evidence ledger. Reviews should report verification differences instead of repeating stable facts.
6. Scale independent verification depth by risk: exhaustive checks for core conclusions, integrity, security, and asynchronous terminal state; sampling or structural checks for low-risk documentation.
7. Generate the next-round contract automatically from review findings and limit each round to one main objective with a small number of acceptance items.

Expected benefits

These changes should reduce empty review cycles, lower retry cost, preserve implementation autonomy, limit final-round expansion, and retain strong reproducibility and auditability.