| name | mm-final-verify |
| description | 数学建模竞赛 V2.6 最终验收阶段。用于检查完整产物、图表插入、结果追踪、论文编译、代码可复现、高分论文评分和提交就绪状态。 |
| allowed-tools | Bash(*), Read, Write, Edit, Grep, Glob, Agent, WebSearch, WebFetch |
Final Verify
文件关系全貌请见 [[FILE_RELATIONSHIP_MAP]] · 上游: [[skills/mm-revision-integrator/SKILL|Phase 5 Revise]](或 [[skills/mm-contest-review/SKILL|Phase 4 Review]] 若无需修订) · 下游: 无(终点) · 共享规范: [[skills/_references/SKILL|_references]]
Load First
Read:
../_references/v2_pipeline_contract.md
../_references/contest_score_rubric.md
../_references/rag_usage_contract.md
../_references/source_quality_policy.md
../_references/anti_template_review.md
../_references/figure_evidence_map.md
../_references/evaluator_optimizer_protocol.md
../_references/executable_model_templates.md
../_references/figure_quality_standard.md
../_references/nature_figure_integration_guide.md when optional nature-figure scientific plotting integration was used
../_references/agent_review_protocol.md
../_references/ars_v2_integration_guide.md when optional ARS integrity checks are available
Required Output
Write reports/VERIFY_REPORT.md with PASS, CONDITIONAL_PASS, or FAIL.
Before writing the final conclusion, run or replicate this read-only audit:
python ../_references/scripts/audit_v2_run.py --workspace <contest-workspace>
If the script path differs because the skill is copied into another root, resolve it relative to the current skills/_references/scripts/ directory. Copy every BLOCKER or HIGH finding into VERIFY_REPORT.md and reports/REVISION_ACTIONS.md; do not return PASS while such findings remain unresolved.
Verification Checklist
Check:
- required V2 files exist
reports/AGENT_RUNS.md exists when subagents or simulated subagents were used
- each
reports/AGENT_RUNS.md entry records goal, input artifacts, model/reasoning, permission scope, output artifacts, conclusion, and thread/id when native threads are available
reports/HUMAN_MODEL_REVIEW.md records final route approval
reports/HUMAN_MODEL_REVIEW.md, reports/MODELING_DECISION.md, and WORKFLOW_STATE.md do not contradict each other
results/RESULTS_MANIFEST.json has metrics and figures
- if any
code_templates RAG hit was used, reports/TEMPLATE_ADAPTATION_LOG.md exists and maps current fields, kept/deleted metrics, applicability, and residual template names
reports/REFINEMENT_LOG.md exists when evaluator-optimizer loops were run, and max-round residual HIGH/BLOCKER issues are copied to REVISION_ACTIONS.md
reports/METHOD_IMPLEMENTATION_MATRIX.md exists and has no unresolved not_implemented core method rows
reports/FIGURE_AUDIT.md exists and inserted paper figures have no FAIL status
- every core figure in
reports/FIGURE_PLAN.md binds to a paper claim, source data, required metrics, axes/units, and an entry in figure_evidence_map.md
- if
nature-figure was used or available for core paper figures, core paper figures have figure contracts, selected-backend scripts, traceable source data, SVG/PDF vector exports when feasible, preview exports, and conclusion-forward captions
- no core data figure relies on
Pillow as its Nature backend
reports/FIGURE_AUDIT.md uses the extended figure-evidence columns when Nature rules are enabled
- every paper figure path resolves
- LaTeX contains
\includegraphics or Typst contains #figure( / image(
reports/CLAIM_TRACE.md has no missing core claims and no strongly worded weak core claims
reports/PAPER_SCORECARD.md exists and all critical dimensions are acceptable
reports/PAPER_SCORECARD.md records judge-skim and anti-template findings when those reviews were run
- core RAG evidence cited in
MODEL_CANDIDATES.md, PAPER_BUILD_REPORT.md, or final claims is S/A; B is auxiliary only and C/D is risk-only
reports/PROJECT_RETROSPECTIVE.md exists at project close, or VERIFY_REPORT.md explicitly states why retrospective distillation was deferred
reports/REVISION_ACTIONS.md and reports/REVISION_STATUS.md show no unresolved BLOCKER or HIGH items
- code files compile or have a documented execution method
- paper compiles where the engine is installed
- no final paper text exposes internal workflow file names, agent logs, or unfinished placeholders
- Optional ARS reference integrity check: if ARS is available, load
<ARS_ROOT>/academic-pipeline/agents/integrity_verification_agent.md as a role prompt. Use only for reference existence, bibliographic accuracy, citation context, and figure/table caption fidelity. Record unverifiable or inaccurate references in VERIFY_REPORT.md; convert actionable issues into REVISION_ACTIONS.md.
- Optional ARS claim faithfulness check: if ARS is available, load
<ARS_ROOT>/academic-pipeline/agents/claim_ref_alignment_audit_agent.md as a role prompt, adapted to V2 evidence. Check claim-to-evidence faithfulness against CLAIM_TRACE.md, RESULTS_MANIFEST.json, figures, code outputs, and METHOD_IMPLEMENTATION_MATRIX.md; do not require ARS citation anchors. Rate claims SUPPORTED, UNSUPPORTED, AMBIGUOUS, or RETRIEVAL_FAILED, and mark strong conclusions on weak evidence.
If native Codex subagents were unavailable and roles were simulated, reports/AGENT_RUNS.md must explicitly say simulated and still record the same fields except native thread/id.
Hard Gates
Return FAIL if any condition is true:
- no figures are inserted into the paper
- generated images exist but are unused without justification
- final paper lacks model validation or sensitivity/robustness analysis
- core numerical conclusions are not traceable to results
- paper cannot be found
- human modeling review is missing
- human modeling review still says the route is waiting/unconfirmed while
MODELING_DECISION.md claims approval
- subagent or simulated-subagent work is claimed but not logged in
reports/AGENT_RUNS.md
METHOD_IMPLEMENTATION_MATRIX.md is missing or contains not_implemented for an approved core method
- the paper claims a method, model, validation, or diagram that code/results/reports did not implement
FIGURE_AUDIT.md is missing, or any inserted figure has failed readability, broken path, garbled labels, square-box text, or unreadable axes
PAPER_SCORECARD.md has any critical dimension below 4 without an accepted contest-specific justification
REVISION_ACTIONS.md contains unresolved BLOCKER or HIGH items, or REVISION_STATUS.md is missing after such actions were created
- a core claim is
missing, or a weak core claim is written as a strong conclusion in the final paper
- generated core figures are unused without a documented paper/appendix reason
nature-figure was used but a core inserted figure lacks selected-backend provenance, source-data traceability, required vector export, or a resolved figure contract
- Nature is available and a core paper data figure is PNG-only, generated with
Pillow, or missing SVG/PDF export without a documented raster-only exception
scripts/audit_v2_run.py reports any BLOCKER or HIGH issue
- a core modeling route or paper claim relies on
B/C/D RAG evidence as its main evidence
- code templates were used but
reports/TEMPLATE_ADAPTATION_LOG.md is missing or leaves unresolved template field names, template paths, or template figure names
- a core figure lacks claim binding, source data, required metrics, axes/units, or evidence-map support
- a formal contest paper with four or more subproblems is shorter than 8 pages without a documented short-report mode or appendix evidence
- a formal contest paper with four or more subproblems is 5 pages or fewer and lacks sufficient formulas, result tables/figures, or validation evidence
Return CONDITIONAL_PASS only when all hard gates pass and remaining issues are limited to MEDIUM/LOW formatting, wording, appendix coverage, or environment issues that do not affect mathematical correctness, evidence traceability, model honesty, or figure readability.
Return PASS only when:
- all contest score dimensions are at least 4, or any exception is explicitly justified and accepted in
VERIFY_REPORT.md
- no unresolved
BLOCKER or HIGH action exists
- inserted figures pass visual/readability audit
- approved methods are implemented or honestly downgraded
- core claims are traceable and worded at the strength supported by evidence
- the PDF or final document has been opened/rendered or otherwise visually checked enough to detect blank pages, garbled figures, and layout failures
Report Format
# Verify Report
结论:PASS / CONDITIONAL_PASS / FAIL
## Evidence Checked
## Hard Gate Results
## File And Figure Integrity
## Claim Trace Results
## Method Implementation Results
## Figure Audit Results
## Automated Audit Results
## Score And Revision Gate Results
## Reproducibility Results
## Paper Build Results
## Required Fixes
Do not call the project complete unless the report conclusion is PASS.