mm-revision-integrator

name	mm-revision-integrator
description	Mathematical modeling contest V2.6 revision loop. Use after mm-contest-review when PAPER_SCORECARD or REVISION_ACTIONS contains BLOCKER, HIGH, MEDIUM, weak-claim, figure-audit, or method-implementation issues that must be repaired before final verification.
allowed-tools	Bash(*), Read, Write, Edit, Grep, Glob, Agent, WebSearch, WebFetch

Revision Integrator

文件关系全貌请见 [[FILE_RELATIONSHIP_MAP]] · 上游: [[skills/mm-contest-review/SKILL|Phase 4 Review]] · 下游: [[skills/mm-final-verify/SKILL|Phase 6 Verify]]（或回到 Phase 4） · 共享规范: [[skills/_references/SKILL|_references]]

Load First

Read:

• ../_references/v2_pipeline_contract.md
• ../_references/contest_score_rubric.md
• ../_references/evaluator_optimizer_protocol.md
• ../_references/figure_quality_standard.md
• ../_references/nature_figure_integration_guide.md when optional nature-figure figure actions exist
• ../_references/agent_review_protocol.md
• ../_references/ars_v2_integration_guide.md when ARS-origin review findings exist

Inputs

Require:

• reports/PAPER_SCORECARD.md
• reports/REVISION_ACTIONS.md
• reports/METHOD_IMPLEMENTATION_MATRIX.md
• reports/FIGURE_AUDIT.md
• reports/CLAIM_TRACE.md
• paper/
• results/RESULTS_MANIFEST.json

Required Outputs

• revised paper, figures, code, reports, or manifest entries as needed
• reports/REVISION_STATUS.md
• updated reports/CLAIM_TRACE.md
• updated reports/METHOD_IMPLEMENTATION_MATRIX.md
• updated reports/FIGURE_AUDIT.md
• updated WORKFLOW_STATE.md

Procedure

1. Read every BLOCKER, HIGH, and MEDIUM item in reports/REVISION_ACTIONS.md.
2. Repair BLOCKER and HIGH items before final verification. Do not downgrade severity just to pass.
3. Repair MEDIUM items when they affect clarity, figure usefulness, method credibility, or judge objections. Otherwise carry them forward with explicit justification.
4. If a promised method was not implemented, either implement it or revise MODELING_DECISION.md, METHOD_IMPLEMENTATION_MATRIX.md, and the paper wording so the route is honest.
5. If a core claim is weak, either strengthen evidence or rewrite the paper claim as exploratory, limited, or conditional.
6. If inserted figures fail visual audit, regenerate or replace them before final verification. 6a. For nature-figure-origin, automated audit, or publication-figure actions, regenerate through the selected backend only, update SVG/PDF/preview exports, refresh source-data references in RESULTS_MANIFEST.json, and update the related figure contract in FIGURE_PLAN.md. 6b. If an action reports Pillow data figures, PNG-only core figures, missing vector export, missing selected-backend script, missing source-data trace, or incomplete FIGURE_AUDIT.md columns, repair the artifact and rerun or replicate ../_references/scripts/audit_v2_run.py --workspace <contest-workspace> before marking it resolved.
7. For ARS-origin findings, treat ARS as reviewer evidence and priority guidance only. Do not let ARS directly override V2 artifacts; apply fixes through the owning V2 stage and record the result in REVISION_STATUS.md.
8. Update REVISION_STATUS.md with one row per action item.
9. For resolved BLOCKER/HIGH actions, optionally log a revision_resolved event with python scripts/memory_log.py from the repository root, or ensure memory_distill.py can later extract the lesson from REVISION_STATUS.md and REVISION_ACTIONS.md.

Revision Status Format

# Revision Status

| Action ID | Severity | Fix Applied | Files Changed | Verification Evidence | Status |
| --- | --- | --- | --- | --- | --- |

Status must be resolved, carried_with_justification, or unresolved.

Final verification may not return PASS while any BLOCKER or HIGH item is unresolved or carried_with_justification.

BLOCKER or HIGH actions from automated audit may not be marked carried_with_justification unless the user explicitly selected short-report mode or the artifact is documented as a non-paper/non-core figure. Otherwise they must be repaired or the final status remains CONDITIONAL_PASS/FAIL.