write-model-paper

Write Model Paper

Reconcile every abstract/body/caption/table number with the same saved value and denominator. Review figures at their final inserted size on the rendered manuscript pages; a standalone export cannot establish manuscript readability.

Read references/practical-figure-insertion.md for specialized figures. Use saved values and denominators in captions and prose; retain one semantic palette across the paper. Internal evidence IDs can remain in saved registries without visible [claim:ID] markers. Explicit venue controls layout; profile controls process intensity.

Write the paper a contest reader sees: mathematical reasoning, evidence, figures and readable pages. Evidence checks protect the draft backstage; they must not turn the paper into an engineering log.

Workflow

1. Confirm the contest template, problem source, model-selection rationale, results, figures, validation evidence and citations. Do not draft numbers, references or claims that lack a source.
2. Read references/paper-generation-guide.md and references/paper-structure.md. If the working project provides config/paper_style.yaml, treat it as an optional local override rather than a required file. For CUMCM 2026, also read references/cumcm-2026-authoring.md. Use current official rules over repository defaults.
3. Draft an abstract that fits one page: problem essence and overall framework, then each question's method, decisive result and necessary validation; end with a short scope statement and 4--7 keywords.
4. Build the paper in the canonical order: abstract, keywords, restatement, analysis, assumptions, notation, conditional unified model, each question's model/solution/results, validation and sensitivity, evaluation, AI-tool declaration when required by the venue, references and appendix. CUMCM papers do not include a contents page.
5. Add a unified-model section only when upstream dependency analysis establishes a shared core. Otherwise make the separate routes clear rather than forcing a formal common layer.
6. For every formula, write purpose before it and definitions, computation and validity after it. For every result, write result, pattern, mechanism, evidence and implication; include limitations where they affect a decision. Read references/evidence-and-consistency.md when drafting results, revising claims or reconciling repeated numbers. Distinguish a supported mechanism from a proposed explanation, and keep terminology, precision and uncertainty definitions synchronized across prose, figures and tables.
7. Use figures and three-line tables as evidence, not decoration. Put a table title above the table; retain representative rows in the body and move exhaustive data to an attachment.
8. Cite only sources actually needed for methods, non-original theorems or external facts. Put references and appendix on separate pages. The appendix must list every support-package file and include every complete, self-authored runnable source file at readable size; list third-party dependencies and installation steps instead of embedding library source or binaries.
9. Render and visually inspect the PDF. Check that the abstract is electronic page 1, restatement follows it with continuous numbering, no contents or blank contents page exists, venue-specific page and package limits (30 pages and 20 MB (local cap: 20,000,000 bytes) for the configured CUMCM baseline) are respected, and no page, table, figure, bibliography or code is clipped or garbled.
10. Build the support ZIP/RAR from a manifest, verify the manifest exactly matches archive entries, include all local modules/data/config/fonts/plot assets, and rerun the declared reproduction command from a clean extracted temporary directory. RAR validation requires rarfile and a working unrar, 7zip, unar, or bsdtar backend. Any missing dependency or nonzero run blocks submission readiness.
11. For CUMCM 2026, place AI工具使用声明 before references. If AI was used, generate AI工具使用详情.pdf from actual records (use assets/ai-tool-usage-details-template.md), include it in the support manifest/archive and reconcile every adopted non-language output with a manual verification record.