chengziyue1222/math-model-agent/run-modeling-project

Run Modeling Project

Coordinate the existing Skills around a strong final paper. The project record supports reproducibility, but it is not a second product or a substitute for modeling.

Workflow

1. Read references/workflow-profiles.md and choose rapid, competition (the default), or audit before creating deliverables.
2. Read references/distilled-competition-playbook.md, then record the problem, data sources, deliverables and venue constraints. Apply CUMCM-only layout rules only when that venue requires them.
3. Invoke analyze-model-data where data is used and select-model for decomposition, dependency analysis and structure opportunities. Classify each needed input as observed, externally supplied, model-derived, or unavailable. Before blocking on a missing derived input, decide whether the problem expects it to be estimated from information available at the decision time; if so, make that estimation a leakage-safe upstream subproblem. Invoke research-model-literature when external facts, parameters, datasets, or citations affect a claim.
4. Invoke solve-model for equations, algorithms, baselines, constraints, sensitivity and independent checks.
5. Invoke make-model-figures only after identifying the conclusions each figure must establish.
6. Invoke write-model-paper to turn verified results into a contest paper, then review-model-paper for reader-visible and technical review.
7. In competition, produce a formal PDF by default (PDF or DOCX is sufficient), a complete support archive and a passing clean-directory/submission preflight. In audit, add both document formats, independent review, HMAC and complete hash/producer evidence.

Priority Order

1. final paper quality and layout;
2. modeling and reasoning;
3. validation and consistency;
4. engineering convenience.

Read references/paper-first-flow.md for the actual handoff order. Existing manifests, stage gates and validations are permitted as backstage controls; their terms and traces do not belong in the submitted paper. In rapid, retain the non-negotiables but defer audit-only artifacts; do not label the result as formally verified.

Handoff Minimums

• analysis → selection: question decomposition, data/units, constraints and dependencies;
• selection → solution: model rationale, rejected alternatives, baseline and validation plan;
• solution → figures/writing: results, formulas, source tables, validations and limitations;
• figures → writing: claim, source, figure role, final artifact and interpretation;
• writing → review: manuscript, selected rendered paper format, complete support manifest/archive, clean reproduction report, sources and a list of evidence used.

A handoff is a reviewable hypothesis, not an irreversible instruction. If downstream data, solver behavior or validation contradicts an earlier choice, return to the affected stage, record the evidence and reason for the change, then regenerate dependent artifacts.

Use distinct status fields for the task result, guidance application, executable-contract runtime, and submission readiness. Never summarize a direct script run as a verified Skill run when the contract runtime was unavailable. Keep one current project state; mark older reports SUPERSEDED with a pointer to the replacing state instead of leaving contradictory current-looking reports.

Resources

• references/distilled-competition-playbook.md — evidence-first workflow distilled from staged competition materials, including rejected shortcuts.
• references/paper-first-flow.md — paper-facing handoff order.
• references/stage-gates.md — backstage stage gates.

Executable Contract

For repository-managed work, inspect the selected profile with python <skill-directory>/scripts/_runtime/skill_contracts.py --skill run-modeling-project --profile competition (or rapid/audit) and pass the same --profile to scripts/execute_skill.py. Rapid does not require the complete role contract; competition uses the delivery contract without audit-only signatures/registries; audit retains the complete signed contract. Record a required preflight failure as blocked rather than repairing the trace by hand.