Summarize modeling experiment evidence, compare the approved main method with a usable baseline, surface fallback triggers, and produce a decision-point or final report without creating routine per-round prose.
Inspect a mathematical-modeling workspace, evaluate lean or submission gates per subquestion, update machine-readable manifests, classify change impact, and route one next action without duplicating downstream work.
Detect whether approved modeling code is Python or MATLAB/Beita Tianyuan and route it to the matching reviewer using the compact named-check review contract.
Audit whether the semantic evidence required by the active lean or submission profile exists and is current, without requiring one verbose artifact per skill or an arbitrary number of pass bullets.
Run scoped or final cross-media consistency checks for mathematical-modeling artifacts, comparing canonical numbers, symbols, parameters, decisions, files, and paper claims without performing full-workspace audits for low-risk changes.
Map contest attachments to subquestions, audit and clean raw data, and emit one reusable data profile with quality, coverage, imbalance, concentration, and method-readiness evidence for downstream risk screening.
Build one compact choice card at a genuine mathematical-modeling judgment point. Use before method screening, after a meaningful experiment, or before final claim/freeze approval so the human chooses the trade-off while AI handles mechanical consequences.
Plan the smallest set of diagnostic, comparison, paper, and appendix figures or tables needed to support verified mathematical-modeling decisions and claims.