Crayonnan/dsh-math-modeling-skills-Gatecraft-

English | 中文

A gated math-modeling skill suite (9 skills + a DSH preset) for DeepSeek Harness. No mindless end-to-end automation — the agent solves and verifies; you think and decide at every stage gate, producing modeling results with your own taste.

What's Inside

• 9 skills: competition-workflow (five-stage pipeline: stage reports / EDA five questions / verification triad) · guozhan-paper (award-paper writing patterns) · vision-ocr (problem & reference-paper reading) · sensitivity-analysis · statistical-diagnosis · math-modeling-paper (paper content) · math-paper-template (LaTeX typesetting) · tex-pdf-image-to-word (convert to Word) · paper-gate (delivery acceptance layer + the single source of the repo-wide constraint-tier convention)
• Loaded on demand: competition-workflow/references/ (model-selection decision tree, pre-submission self-check) and guozhan-paper/references/ (award-paper evidence corpus, page-level samples and counterexamples) stay out of the resident context until needed.
• assets: optimization-playbook (optimization solve/verify decision tables) · figure-playbook (flowchart & figure templates) · prompt-pack (14 battle-tested prompts) · flowchart_gen.py (spec → drawio generator) · ocr_batch.py (concurrent OCR) · docgate.py (paper-gate engine) · official-paper-format.md (official format source)
• DSH preset: presets/math-modeling/ — paste a contest problem and the workflow starts automatically

Install

dsh plugin add Crayonnan/dsh-math-modeling-skills-Gatecraft-

Optional: copy presets/math-modeling/ to ${DSH_HOME:-$HOME}/.dsh/.agent-presets/math-modeling/, then pick "数学建模模式" in a new session.

Usage

Run the competition-workflow pipeline on the problem at [path/attachment].

Flow: read the problem (verify external guides + literature) → data-structure exploration (EDA five questions) → modeling (coherence chain + flowchart spec) → solving (verification triad) → sensitivity / diagnosis → seven-part writing → docgate.py mechanical check + human self-check. Stage reports give every downstream number its source; whether to keep iterating or record a limitation honestly is your call.

Delivery acceptance (paper-gate + docgate.py)

13 mechanical checks over docx (OOXML) and tex, run against the file you actually submit. FAIL is now reserved for objectively wrong facts and truthfulness problems (broken cross-references, abstract numbers absent from the body, figures overflowing the text block, empty variable-explanation residue, restatement too close to the prompt); proportionality, structure shape, repetition and prose style report WARN for human judgment. Severities live in paper-gate-rules.yaml, so changing a contest means editing config, not the script.

Philosophy

• Hard constraints only for facts and truthfulness — official rules, engineering constraints and mechanically verifiable consistency get imperative language, and ideally live in docgate.py rather than prose. Craft judgments (length, sentence form, which test to run) are written as "common pitfall + recommended/not-recommended + self-check question". For a strong model, dressing a craft disagreement as a blocking condition doesn't make it stricter; it just invites gaming the metric.
• Stage reports are not approval gates — their job is to give the next stage's numbers a source. 2-3 iterations per sub-question is usually enough; log "change → effect → metric" each round.
• Report first: every sentence in the paper is derived from facts in the stage reports; sample sentences are never copied
• Number discipline: every number traces to a report or code output; zero-drift re-check after reruns
• Critical verification: verify external guides item by item, recompute third-party claims, benchmark results against literature
• Taste from patterns: the four coherence requirements (R1-R4) each carry "criterion + positive sample (with page) + counterexample (with page)"; award-paper sentences serve as samples, not templates

Vision (for text-only agents)

Default channel: SiliconFlow Qwen3-VL (API key required via env var SILICONFLOW_API_KEY, sign up at cloud.siliconflow.cn/me/models). No key is bundled with the plugin — set the env var before use. Local alternative: qwen-mm-plugins. Or point ocr_batch.py at your own OpenAI-compatible vision model via BASE_URL/MODEL. Without any vision channel, flowcharts still pass QA through the "spec → drawio/PNG → OCR read-back" loop (figure-playbook §4).

Scope

Battle-tested on statistical-analysis and optimization/decision problems (the typical "C" problem). Mechanism/physics-simulation (A) and graph/engineering (B) problems are untested — extend the checklists yourself and contribute back.

With MathModelAgent

Division of labor, not duplication: its solvers serve as a backend (mma_exec_python hooks are pre-reserved), GateCraft is the orchestration & QC layer — thinking, pivoting and deep participation happen at the stage gates.

Layout

skills/         9 skills (competition-workflow is the orchestrator)
assets/         playbooks / prompt-pack / docgate.py / generators (synced with skills)
presets/        math-modeling (DSH preset)
index.js + cordis.patch.yml + package.json   dsh bundle packaging

Creation Story (expand)

Origin. The 2023 CUMCM paper C228 (national first prize) shows "coherence" comes from four mechanisms: positioning statements, model-choice motivation chains, the three-part result explanation, and explicit reuse declarations. Our 2026 Huashu Cup C paper proved solving depth can clear the prize bar while figure cross-references go wrong, transition paragraphs go missing, and scope clauses never enter the body. The TipDM Cup C and Greater Bay Area Cup B papers (two second prizes) complete the defect list with seven classes: abstract-body number drift (the "1.87%/99.2%" figures exist nowhere in the body), internally inconsistent table columns, 29 mixed "图表N" captions, leftover "[GPT-5, OpenAI]" annotations, entropy weighting over n=2, AUC reported on 5 positive samples, and misused BH-FDR. The 2025 paper C023 (national first prize, later journal-published) sets the benchmark: seven-part structure per question, two-layer flowcharts, diagnose-before-modeling. GateCraft is the solidification of all these lessons.

Round 1: C228's coherence → the R1-R4 requirements + rigor language rules; added stage 0.5 data-structure exploration (EDA five questions, each finding tagged "→ which model design it decides"). A smoke test on the real dataset caught two scope errors in our own paper on the spot.

Round 2: the seven defect classes → abstract three-way reconciliation / table-column self-consistency / AI-trace scanning / method-sample-size matching; C023 → seven-part structure + two-layer flowchart spec. Second-hand analyses were verified item by item against primary sources ("PSO grouping", "figures 1-1~1-5", "21.3/28.6" all falsified). OCR accelerated from serial 32B to 8B default + 32B recheck + 4-thread concurrency: 84 pages in ~12 minutes (~6×).

Round 3: two full session logs → 14 battle-tested prompts distilled into prompt-pack (each with "when / template / criterion / measured effect").

Gotchas: the flowchart QA loop (spec JSON → generator → OCR read-back; κ→k needs Greek-Latin normalization); the DSH preset mount validation discovered tool-cordis registers process-global providers — two cordis-family presets cannot coexist in one process, so the preset ships without self-modification tools; the open-source audit separates the private repo from the release snapshot, which passes "no keys / no personal paths / no paper extracts" checks before archiving.

License

MIT. Contributions follow one format: requirement / decidable criterion / positive sample (with page) / counterexample (with page) — every checklist item must come from a real failure or a real award.