review-model-paper

Review Model Paper

Report mathematical evidence, figure expression and page-layout defects separately in the existing review record. After a repair, inspect the regenerated dependent outputs and their current hashes; keep unresolved uncertainty and unviewed pages pending.

For practical matrix/calendar/composition figures or a failed stage gate, read references/practical-figure-review.md. Check source numbers, units, composition denominators, size/height/angle encoding, missing states and final insertion. Return the earliest affected stage; release requires passing review, preflight and current evidence, without fixed counts of equations, figures or citations.

Review in read-only mode unless revision is explicitly requested. First judge the reader-visible paper, then run backend consistency checks.

Workflow

1. Collect the problem, manuscript, selected paper delivery (PDF by default; PDF or DOCX for competition, both for audit), result files, figures/tables, validation material, citations, support archive/manifest and contest template.
2. Read references/paper-quality-review.md and references/review-rubric.md. For CUMCM 2026, also read references/cumcm-2026-preflight.md and label each finding as official compliance or quality advice.
3. Check abstract, problem analysis, unified-core decision, formula explanation, per-question model/solution/result coverage, result reasoning and limitation disclosure.
4. Check that figures and tables have different evidentiary roles, are readable on A4 pages, and receive useful surrounding prose.
5. Check citations for relevance and verifiability. Require the appendix to list every support-package file and contain every complete self-authored runnable source file; third-party libraries belong in a versioned dependency list with installation instructions, not as copied source or binaries.
6. Render and inspect layout against the selected venue/template. For CUMCM, the abstract is electronic page 1, the restatement follows with continuous numbering, and no contents page exists. References and appendices must read cleanly; the body obeys the resolved venue page limit. Treat typography and heading alignment as local quality defaults unless explicitly mandated.
7. Keep ordinary review read-only. For a final anonymity preflight, use a user-supplied private identity dictionary outside the support tree or explicit --identity-term values when available. Do not create or populate an identity file during review. If no private dictionary is supplied, still run generic identity-label, path, filename and metadata checks and report the personalized term scan as not run rather than failing the whole review.
8. Run structural, value-consistency, resolved venue package-size/anonymity, path and PDF/DOCX metadata checks. Verify the support manifest exactly matches ZIP/RAR entries and rerun it from a clean extracted temporary directory. For CUMCM 2026, require the AI declaration before references and, when AI was used, the exact support file AI工具使用详情.pdf. Reconcile disclosure against records rather than guessing from writing style. RAR review requires rarfile plus a detected extraction backend and records that backend in the report. Any missing local import/data/config/font/plot dependency, confirmed identity leak or absolute user path blocks readiness.
9. Report findings by severity with direct evidence, impact and the smallest repair. A PASS means the visible paper and its evidence both meet the declared standard, not that it will win a contest. Use references/evidence-review.md for claim/source pointers, statistical ambiguity, figure semantics and a revision closure check. Separate confirmed defects, unresolved evidence gaps and optional aesthetic preferences.