At a judgment point, emit the 2-3 questions only the human modeler can answer — framed as trade-offs, not answers — and refuse to answer them. The inverse of "AI answers, human confirms": here the AI asks, the human answers, then the AI assists with the consequences.
Manage the full mathematical modeling contest workflow and decide which skill should be used next.
Verify every skill that claims "completed" produced a substantive audit/review artifact on disk with ≥ 5 explicit pass items. Runs as part of the independent audit layer that does not trust any single skill's self-declaration of done.
Plan figures and tables that support the modeling logic, results, and paper narrative.
Generate publication-quality mathematical modeling figures with matplotlib, covering evaluation charts, prediction plots, optimization diagrams, mechanism schematics, heatmaps, and multi-panel layouts. Use when creating or revising figures for contest papers.
Extract, organize, and document unified model assumptions from the problem parse and candidate method pools, distinguishing necessary from simplifying assumptions.
Write mathematical modeling paper sections based only on available problem analysis, model plans, results, figures, and robustness reports.
Classify each subquestion into standard mathematical modeling problem types.