Lucasuiii/modeling-workbench / .agents/skills/cumcm-workflow
cumcm-workflow
VerifiedBuild or resume mathematical-modeling competition work, including CUMCM, MCM/ICM, graduate and regional contests, data challenges, and open-topic statistical modeling. Guide problem framing, modeling, computation, validation, Chinese or English LaTeX writing, and reviewed PDF/source delivery using current official requirements. Not for ordinary paper polishing.
references/06-paper-writing.md
View on GitHub# Paper responsibility
Start in a fresh task when practical. Read `handoffs/validation-paper/HANDOFF.json` first; open upstream files only through its canonical pointers. Do not load failed runs, debug logs, or old review conversations unless a current P0 specifically requires them.
Treat handoff `representation_candidates` as prompts for judgment, not a prescribed chart list. They flag evidence shapes that may benefit from visualization even when no figure exists yet. Select prose, equation, table, or figure according to the claim. Handoff limitations contain claim limits, P1 concerns, and model applicability/assumption/known-limit information; do not reinterpret model scope itself as a limitation.
For English work, audience-specific documents or reference verification, read [competition writing](competition-writing.md). Follow current official requirements while preserving the reviewed claim scope. Set the actual competition and paper language when initializing the scaffold; see [competition adaptation](competition-adaptation.md). Both languages use the same compile, page-review and delivery chain, with official templates taking priority.
## Plan the argument before LaTeX
Default to a fully developed, visually informative paper. Give important reasoning and results enough space for a reader to understand and assess them without opening the code or reconstructing missing steps. A short conversational answer is not the target form of a competition paper. Respect the user's requested depth and the current official page limits; spend space on the decisions and evidence that matter most.
1. Select the validated claims that answer the official questions.
2. Actively look for useful visual explanations in the handoff candidates and existing evidence; choose complementary prose, equations, tables and figures for the important claims.
3. Design `paper_structure` as the semantic argument flow.
4. Generate the selected tables/figures from existing results and draw helpful model/algorithm diagrams. Do not stop at whatever images the computation stage happened to export.
5. Adapt a declared official paper template when supplied; otherwise initialize the generic scaffold and keep compliance unverified until rule/instruction documents are checked.
6. Write, compile, render every page, and review the PDF.
`paper_structure` is the source of truth for the body. Each entry supplies a section title, purpose, covered subproblems, and supported claims. One section may serve several subproblems; one complex subproblem may span several mathematically meaningful sections. Background, restatement, assumptions, notation, shared mechanisms, model development, results, validation, evaluation, and conclusions are candidate modules—not mandatory headings. Cover every official subproblem, but do not manufacture sections to imitate a generic modeling paper.
Sections are not independent essays stapled together. Each one states, in its opening sentence or two, why the previous section made it necessary: a quantity the previous model left unknown, an approximation that has to be checked, a decision the previous result did not settle. Questions that are formally independent are still tied together by the shared object, data, or mechanism they act on. A reader should be able to say what breaks if a section is removed. This costs a sentence per section and it is the difference between an argument and a list.
`PAPER_PLAN.json` needs `claim_selection`, `representation_plan`, `paper_structure`, and `authoring_task_ref` — the task writing the paper, which must differ from the `validation-paper` handoff's `producing_task_ref` (see `references/handoffs.md`). Legacy argument layers, reference-paper counts, page budgets, and figure counts may remain optional notes but are not hard gates. A plan with no table/figure creates a warning to reconsider communication, not a failure.
## Develop the explanation
For the central models and results, connect the question, modeling choice, assumptions, formulation, solution and interpretation in continuous prose. Explain why a key objective or constraint represents the real problem, how a non-obvious equation follows, and what each important variable and parameter means, including units. Describe enough of the actual algorithm, stopping criteria and parameter selection for a reader to understand the computation. Allocate detail according to mathematical difficulty and consequence; a familiar algebraic step needs less space than a consequential approximation.
After reporting a result, explain its magnitude, pattern, comparison and practical implication using the available evidence. Describe where performance changes or a constraint becomes active, and connect that observation to the model when justified. Distinguish an observed pattern from a plausible explanation when the mechanism has not been tested. Avoid leaving the reader with a formula, a table and a sentence saying that the method works.
Develop short technical notes into coherent paragraphs where the reasoning needs it. Keep lists and tables for genuinely parallel items, avoid repeating the problem statement as background, and move lengthy code or routine derivations to an appendix when permitted. Under a page limit, trim repetition and supporting bulk before removing the explanation needed to assess the main answer. These are writing defaults, not mandatory paragraph templates or word counts.
## Abstract and keywords
Write the abstract after the body stabilizes. Use the order problem -> core method -> key result -> meaning/validation. For a quantitative task, include a few decision-bearing numerical anchors with units, comparison, uncertainty, or fit information when supported. Wrap the values that ARE the answer in `\keyresult{}` so a reader skimming the abstract finds them without parsing the sentences around them; use it in the body's result statements too. Bold the answer, not every number on the page — marking everything marks nothing. Avoid empty sequences such as “a model is built for Question 1; Question 2 is solved; results show effectiveness.”
Keywords must come from the actual object, data, model, or method. Do not use workflow filler such as “mathematical modeling,” “reproducible computation,” or “evidence chain.”
## Claim-serving representations
Actively favor a useful additional figure when it reveals a pattern, mechanism or comparison that would otherwise take substantial prose to explain. Inspect the actual indexed outputs behind promising handoff candidates; an empty figure directory does not mean there is nothing to visualize. Record selected representations in the existing `representation_plan`, with their purpose and evidence, rather than creating another planning artifact.
Every representation answers one reading or evidence question:
- data-coverage or distribution views explain the observations and their limits when the corresponding summaries already exist;
- observed-versus-fitted plots show where the model follows or misses the data;
- residual/error plots expose structure hidden by a global fit score;
- comparison plots support method, scenario, or policy choice;
- sensitivity/convergence plots support parameter or algorithm stability;
- robustness distributions support noise or perturbation claims;
- mechanism/algorithm diagrams clarify a genuinely complex process. These are drawn, not plotted: an optical path, a four-stage decision flow with a rework loop, a state transition, an algorithm's stages. They need no data, which is exactly why they get skipped — an author reaching for matplotlib finds nothing to plot and moves on. The scaffold loads TikZ so drawing one costs nothing;
- compact tables carry values that readers must compare or retrieve precisely.
These are options, not a checklist. Do not require every kind, impose a minimum count, or invent residual, Monte Carlo, sensitivity, convergence, or robustness analyses that computation/validation did not execute. The initializer never decides what to plot.
Use complementary views when they answer different questions: a prediction curve can locate errors, an existing error distribution can reveal their spread, and a compact metric table can give exact comparisons. Avoid several chart types repeating the same observation. Draw mechanism/algorithm diagrams from the actual model and distinguish schematic illustrations from measured results; decorative images do not improve an argument.
Introduce each substantive figure with the question it helps answer, then discuss its important feature, evidence-supported interpretation and consequence in the surrounding text. Captions identify the data or scenario, axes/units, groups and uncertainty where applicable; the body explains why the reader should care. Keep plotting code and result bindings in the existing figure/evidence path. Replotting established outputs is allowed; a new metric, data aggregation, experiment or stronger conclusion goes back through computation/validation and any affected existing checkpoint before entering the paper.
## Result -> validation -> boundary
After a result, explain what makes it credible, where it weakens, and what changes the conclusion. Prefer already available residual/error analysis, feasibility or constraint checks, sensitivity, convergence, stability/robustness, baseline/model comparison, or out-of-sample evidence. If an important check is missing, state a concern or limitation; the paper stage must not create a new numerical experiment merely to complete the narrative.
Distinguish the two kinds of cross-check and say which one you have. Computing the same quantity twice through the same model — a closed form against a truncated series, a vectorized implementation against a loop — agrees to machine precision and shows only that the code matches the derivation. Computing it through a second, independent route — a different physical principle, a different estimator, a different data channel — produces a real discrepancy that has to be explained, and that explanation is what convinces a reader the model is right. Implementation agreement is the floor. Do not present it as model validation.
When model evaluation is useful, distinguish strengths, limitations, and possible improvements. Tie each point to the actual mechanism, assumption, data region, parameter identifiability issue, sensitive perturbation, or missing data. Prefer claim limitations and accepted/open P1 concerns over generic claims that a model is simple, accurate, general, or practically meaningful.
## Quality-reference boundary
The user-supplied 2025 B first-prize paper, *基于联合物理色散模型与先进信号处理方法测定碳化硅外延层厚度*, is a quality reference, not a template.
Transferable lessons:
- a high-information abstract names the physical/modeling route and reports numerical anchors;
- derivation, algorithm, computed result, interpretation, and reliability analysis form a continuous argument;
- fitted curves, residuals, method comparisons, sensitivity/convergence views, and robustness distributions have distinct claim functions;
- limitations identify concrete systematic residuals, material assumptions, parameter coupling, and dependence on external constants;
- long supporting code and file lists remain in the appendix instead of interrupting the main narrative.
Non-transferable traits:
- its section count/order, 38-page length, and number of figures/tables;
- its optics models, algorithms, parameter values, and physics-specific organization;
- its AI-tool citation practice and any 2025 competition-specific formatting;
- any claim that another paper should look similar or reproduce analyses absent from current evidence.
Use this priority whenever sources conflict: a current official paper template and official rules > the generic scaffold > reference-paper style. Rule PDFs/DOCs are compliance inputs, not automatically adaptable templates.
## Reader-facing quality
Before finalizing, review the paper as a reader: which key reasoning step is too compressed to follow, which result is merely announced, which existing evidence would be clearer as a figure, and which figure lacks interpretation? Repair specific gaps using the current evidence, and retain additions that improve understanding. Record material findings in the existing quality report; do not add a scoring matrix, minimum length, image quota or extra approval. An optional presentation improvement is P2; a missing explanation that makes a central claim unassessable is classified by its actual impact, not by paper length.
- Explain why equations and algorithms are used and what results mean.
- Prefer a representative result plus interpretation over number dumping.
- Use tables for exact comparisons, figures for patterns/relationships, equations for mechanisms, and prose for conclusions or assumptions.
- Give captions enough context to understand the claim, scope, axes/units, and comparison without repeating the body.
- Keep internal IDs, evidence states, local paths, workflow terms, and run coverage in sidecars only.
- Do not create new numerical results while writing.
- Do not impose a minimum page or chart count.
- Keep figures near the argument they support; use booktabs-style tables, concise captions, and subfigures only when joint comparison improves reading.
- Use `tabularx` or `longtable` for genuinely wide/long material before shrinking text. Move supporting bulk to the appendix when rules allow.
`PAPER_QUALITY_REPORT.json` binds content, rendered layout and final QA to the exact PDF. v0.6 dropped the per-question eight-dimension matrix: a reviewed subproblem records `status` and `notes`, and the layout block's page count, rendered pages and checks come from `record_compile.py` rather than from an attestation. Open P0 issues block final status. P1 concerns and P2 suggestions remain visible but do not block. Failed layout checks that make equations, tables, figures, glyphs or pages unreadable remain hard submission-reliability failures.
Run `paper_visible_text_check.py`. Internal metadata and local paths block; excessive decimal precision and number-dense sentences are warnings for reader review. It matches workflow IDs by their actual shape (`PREFIX-[Qn-]NNN`), so ordinary prose such as "Model-A" is not a false positive. A revision log is optional: use it when it helps track a real P0/P1 correction, not to force rounds.
Build `paper-delivery` after the exact PDF and editable source are approved.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| canvas-designanthropics/skillsCreate beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| docxanthropics/skillsUse this skill whenever the user wants to create, read, edit, or manipulate Word documents (.docx files) or Word templates (.dotx files). Triggers include: any mention of 'Word doc', 'word document', '.docx', '.dotx', or requests to produce professional documents with formatting like tables of contents, headings, page numbers, or letterheads. Also use when extracting or reorganizing content from .docx or .dotx files, inserting or replacing images in documents, performing find-and-replace in Word files, working with tracked changes or comments, or converting content into a polished Word document. If the user asks for a 'report', 'memo', 'letter', 'template', or similar deliverable as a Word or .docx file, use this skill. Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| pdfanthropics/skillsUse this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text/tables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting/decrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable. If the user mentions a .pdf file or asks to produce one, use this skill. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| pptxanthropics/skillsUse this skill any time a .pptx or .potx file is involved in any way — as input, output, or both. This includes: creating slide decks, pitch decks, or presentations; reading, parsing, or extracting text from any .pptx or .potx file (even if the extracted content will be used elsewhere, like in an email or summary); editing, modifying, or updating existing presentations; combining or splitting slide files; working with templates (.potx), layouts, speaker notes, or comments. Trigger whenever the user mentions "deck," "slides," "presentation," or references a .pptx or .potx filename, regardless of what they plan to do with the content afterward. If a .pptx or .potx file needs to be opened, created, or touched, use this skill. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |