Paper Formal Writer
Resolve <SKILL_ROOT> to the parent directory of this loaded skill folder, using its actual absolute path. Run commands from the contest project root, never from the installed skill directory. The placeholder is not a shell environment variable.
This skill is the sole formal author and final manuscript producer in Standard. It turns a fresh S6 evidence chain into an audited Markdown source and native-equation Word document.
Entry
Run before formal work:
bashpython "<SKILL_ROOT>/paper-workflow-orchestrator/scripts/workflow_guard.py" --skill paper-formal-writer
S0-S6 must pass. If workflow state is uncertain, run workflow_guard.py --status and follow its current artifact report rather than memory.
Single Formal Path
textS6 evidence PASS -> writing plan -> chapter drafts -> chapter audits -> queued local repair when required -> deterministic assembly -> Agent global revision -> final Markdown audit -> formal DOCX -> S8 format/render gate
Do not promote quickstart, legacy, or mechanically merged micro-unit output into a formal filename.
Prepare
- Build the evidence-aware outline:
bashpython "<SKILL_ROOT>/paper-formal-writer/scripts/build_paper_outline.py"
- Prepare adaptive authoring:
bashpython "<SKILL_ROOT>/paper-formal-writer/scripts/prepare_authoring.py" --mode auto
Default delivery is competition: plan at least 14000 effective characters, with a main-paper floor of 8000 and at least 18 rendered pages before appendices. These are adjustable project defaults, not contest rules. Verify the actual contest's page cap and counting rules before writing. auto therefore selects section; a short report needs --delivery short-report --scope-reason "user-requested scope". Custom floors use --min-pages / --min-body-chars with a reason; never lower them merely to turn a failed check green.
Read references/competition-scope.md when preparing scope or repairing a short manuscript. Several authoring turns may produce one coherent chapter; preserve evidence and notation between turns instead of compressing the paper into one response.
The authoritative S7 contracts are:
paper_output/plan/writing_plan.jsonpaper_output/context/authoring_state.jsonpaper_output/qa/draft_audit.jsonpaper_output/qa/repair_queue.json
Draft And Audit
For numerical results, read references/metric-claims.md.
Formal preparation binds each required metric to its unit, sample/split, CSV row and
successful experiment. Missing metadata blocks S7 instead of being guessed from prose.
Maintain a sibling .claims.json for each draft and the final source; use an empty
claims list for chapters without numbers. Only declared conversions are allowed.
Never promote a feasible solution to an optimal one without recorded verification.
Write only to the path recorded for the current unit in authoring_state.json. Each chapter must be a complete, coherent section and must include its required removable marker:
html<!-- mathmodel-evidence: evidence-id-1, evidence-id-2 -->
Audit every changed draft:
bashpython "<SKILL_ROOT>/paper-formal-writer/scripts/validate_authoring.py" --section <section-id>
Routing is deterministic:
globalrepeats the same blocking category twice: switch tosection.- A section fails once: rewrite that section.
- A section repeats the same category twice:
paper-micro-unit-generatormay repair only queued locations. - The third changed attempt with the same category: S7 becomes
BLOCKED; report the cause and suggest Lite as a user choice without switching automatically.
Audits block short drafts, missing chapters/evidence/numbers, broken formulas, placeholders, duplicate prose, broken figure/table references, and internal workflow language.
Assemble And Revise
After all active units pass:
bashpython "<SKILL_ROOT>/paper-formal-writer/scripts/assemble_sections.py" python "<SKILL_ROOT>/paper-formal-writer/scripts/validate_authoring.py" --assembled
The assembler follows outline order, strips evidence comments, merges claim sidecars, and preserves existing numbering. It writes paper_output/drafts/assembled_draft.md.
Then read the entire assembled draft and evidence chain, and rewrite the paper globally into:
textpaper_output/final_paper_source.md
This pass must unify terminology, notation, transitions, argument order, citations, captions, and conclusions. Copying the assembled draft unchanged is rejected.
Update final_paper_source.claims.json with the revised quotes and exact source hash;
do not copy stale bindings. Sidecars remain outside the rendered manuscript.
Run the final source audit:
bashpython "<SKILL_ROOT>/paper-formal-writer/scripts/validate_authoring.py" --final
Any change to evidence, writing plan, approved section draft, assembly, or final source invalidates downstream PASS state.
Word And S8
Only after authoring_state.status = PASS:
bashpython "<SKILL_ROOT>/paper-formal-writer/scripts/format_formal_docx.py" python "<SKILL_ROOT>/paper-formal-writer/scripts/check_paper_format.py" --render required
Formal mode requires fresh S6 and S7 hashes. --allow-draft may create final_paper_draft.docx for layout diagnosis, but never satisfies S7/S8 or overwrites the formal DOCX.
format_formal_docx.py converts LaTeX to editable Word OMML and blocks failed conversion. check_paper_format.py checks dynamic length, required hierarchy, citations, evidence-linked figures/tables, formulas, duplicate prose, DOCX structure, LibreOffice PDF rendering, page count, and extractable text.
Writing Standard
- Use
1 / 1.1 / 1.1.1headings and complete CUMCM sections. - For each question, connect assumptions, variables, derivation, algorithm, computed values, validation, uncertainty, and conclusion.
- Define symbols before formulas and explain every displayed result.
- Cite and interpret every included figure/table.
- Use only current evidence; never invent runs, values, sources, or validation.
- Avoid copied or number-swapped paragraphs used only to meet length.
- Keep implementation language out of body prose; reproducibility paths belong in the appendix.
References
- Format standard: references/cumcm-paper-standard.md
- Full-paper structure: references/formal-paper-template.md
- Section expansion: references/section-expansion-rules.md
- Figures, tables, formulas, and results: references/figure-table-writing-rules.md
Read only the reference needed for the current writing or repair decision.
Handoff
After each meaningful transition, run workflow status and update paper_output/context/workflow_memory.json:
bashpython "<SKILL_ROOT>/paper-workflow-orchestrator/scripts/workflow_guard.py" --status python "<SKILL_ROOT>/context-memory-keeper/scripts/update_workflow_memory.py"
S8 is complete only when the machine-readable format report is PASS and fresh.
