竞赛合规与提交

用途

Gate the submission against official rules and reproducibility requirements. Treat compliance as part of the mathematical result: a strong model can still lose if the package is non-compliant, non-anonymous, or unreproducible.

核心规则：官方竞赛规则和用户提供的通知高于本地风格、往年习惯和模板记忆。

国一候选门槛：合规层只放行题目贴合、建模洞察、证据可信、可复现、边界清楚且不触犯规则的提交包；这是提交质量门槛，不承诺获奖。格式漂亮但规则来源、匿名性、AI 披露或复现证据缺失时，必须阻塞最终提交。

硬门禁：只要当前官方规则来源未知、匿名风险未关闭、必需附件缺失、要求披露时 AI 使用披露未解决，或核心代码不能复现论文引用结果，最终提交就必须阻塞。

返回规则：本技能只把发现返回 math-hub。可以建议修复面，但不能直接路由、派发或调用另一个技能。

规则来源门禁

Before final-format or package claims, record:

• contest name, problem id, track, year, and active phase;
• official notice or user-provided rule source, path or URL, and date checked;
• paper page/format requirements, template constraints, naming rules, archive rules, and deadline-sensitive package rules;
• appendix, code, data, support-material, and electronic/paper consistency requirements;
• AI-use rule: allowed, prohibited, disclosure required, disclosure format unknown, or not specified;
• anonymity constraints for title page, headers/footers, PDF metadata, filenames, comments, code, and support materials.

If a rule source is missing or stale, mark final submission blocked; do not infer from previous contests.

合规检查面

Check these surfaces explicitly:

Surface	Blocking risk
Paper PDF	Wrong template, page limit, table-of-contents rule, bad metadata, author identity, unreadable figures
LaTeX/source	Identity in comments, stale template residue, missing files, non-compiling source
Code	No run instructions, missing dependencies, hardcoded local paths, unreproduced cited results
Data/support	Missing required attachments, stale outputs, private or identifying files, unsupported figure/table sources
AI disclosure	Missing usage log, unsupported claim generated by AI, no human verification note
Manifest	Package lacks latest PDF, source, code, data, result tables, figures, logs, or exclusion rationale

Do not package private drafts, prompt logs, unrelated data, caches, or identity-bearing metadata unless the official rules require them.

AI 使用披露门禁

When AI was used or rules require disclosure, ai_usage_log.md must state:

• tool/provider/model or version when known;
• date/time and task phase;
• interaction summary or prompt/response summary sufficient for disclosure;
• generated material that entered paper, code, figures, translation, polishing, verification, or support files;
• human verification and modification note;
• paper, appendix, support-file, or manifest location.

No AI-assisted formal claim, code result, figure, or final text is compliant if it lacks traceable human verification when disclosure is required.

For claim-level AI disclosure, create or audit ai_claim_disclosure.csv when AI-assisted material is adopted into the submitted paper, code, figure package, or support files:

adopted_material_id,claim_id,artifact,location,ai_tool,model_or_version,date,task_phase,interaction_summary,human_verification,modification_summary,disclosure_location,status

规则：

• adopted_material_id must identify the exact adopted text, code, figure idea, translation, verification note, or support-file material.
• claim_id may be empty only for non-claim surfaces such as formatting or translation; final claims need a claim id.
• human_verification must name the concrete check performed, not just "reviewed".
• disclosure_location must point to the paper marker, reference entry, support file, manifest entry, or official disclosure form required by the active rules.
• Generic statements such as "AI was used for assistance" are insufficient when the official rule requires details.

复现门禁

Core cited results must have:

• run command or documented manual procedure;
• entry script, input files, parameters, seeds, solver status, warnings, and output paths;
• run_record.csv row for computed results;
• result_registry.csv and figure_evidence.csv rows when results or figures support claims;
• instructions that work without hidden local absolute paths or personal files;
• latest outputs matching the paper's cited values, filenames, and figure/table numbers.

If the package cannot reproduce key numbers, block submission even if the PDF looks finished.

提交清单约定

Create or audit final_submission_manifest.md with:

Official rule source and date checked:
Paper PDF path and compile command:
Main source path:
Included files and why each is required:
Excluded files and why each is safe to exclude:
Result tables and figures cited by the paper:
Run records that reproduce core results:
Anonymity checks:
AI-use disclosure/support details:
Open P0/P1 risks and accepted P2 risk notes:
Final status:

Final status may be ready, blocked, or diagnostic-only. Use ready only when official rules, anonymity, AI disclosure, package contents, code reproducibility, PDF compile/inspection, open P0/P1 findings are resolved, P2 findings are resolved or explicitly accepted with risk notes, and P3 findings are non-blocking.

输出约定

Return:

Compliance lock:
Official-rule status:
Submission package status:
Anonymity status:
AI disclosure status:
Claim-level AI disclosure:
Code/support reproducibility:
Manifest changes:
P0/P1 blockers:
Smallest repair order:
Return to hub: math-hub
Recommended repair surface:

Recommended repair surface may name latex, code, figure, review, abstract, or hub QC, but the handoff always returns to math-hub.

红线

Stop and block final submission when:

• official page, attachment, AI-use, anonymity, or package rules are unknown;
• identity appears in PDF metadata, filenames, comments, code paths, support files, or appendix text under anonymous rules;
• the paper cites a table, figure, code output, or appendix file absent from the package;
• a required code run depends on local absolute paths, interactive manual steps, or missing data;
• AI use is disclosed generically but not tied to actual adopted material and human verification;
• ai_claim_disclosure.csv is missing required adopted_material_id, claim id, verification, modification, or disclosure location for adopted material;
• final packaging proceeds while open P0/P1 review findings remain;
• the compliance audit starts routing, dispatching, or calling another skill instead of returning findings to math-hub.