Research Review Loop

Quick start

1. Identify the target artifact, stakes, intended audience, and requested evidence class.
2. Initialize review state with scripts/init_review_loop.py unless a review pack already exists. Use --from-paper-review <review_dir-or-final_issues.json> when a first-pass paper-review bundle is available.
3. Build a claim ledger and a predecessor-failure ledger before issuing conclusions.
4. Review for internal consistency, evidence quality, external verifiability, provenance, non-vacuity, complete outcome accounting, and real rather than nominal independence.
5. Preserve previous rounds and append or version new outputs instead of overwriting prior review artifacts.
6. Update the latest REVIEW_STATE.json, AUTO_REVIEW.md, and NARRATIVE_REPORT.md after each round only after preserving prior versions.
7. If a research-paper-review bundle exists, import its issue IDs, quotes, ratings, and summary path rather than re-summarizing or downgrading the first-pass critique.
8. Infer the provenance of existing review artifacts before reorganizing them: upstream Claude/OpenAIReview bundles, Codex-native paper-review bundles, and review-loop hybrids are all valid inputs.
9. If the user asks for project-level sequencing, current-state inspection, or coordination across multiple research stages, invoke research-pipeline-planner first instead of treating the review loop as the whole task.
10. Use ../research-pipeline-planner/references/epistemic-assurance-contract.md whenever a round is expected to authorize evidence promotion or route advancement.

Relationship to sibling skills