zhnnky329/MathModeling-skills

What it solves

This project addresses common workflow failures in mathematical modeling contests, such as misinterpreting problem statements, skipping baseline models, using inconsistent data, or including numbers in a final paper that cannot be traced back to actual code output. It shifts the AI's role from a "ghost-writer" to a technical assistant that ensures mechanical correctness and rigorous auditing while leaving modeling judgments to the human user.

How it works

The system implements a structured pipeline with 28 specialized skills and "hard gates" that prevent progress until specific conditions are met. It operates through several stages:

1. Groundwork: Parses problems, classifies tasks, and builds shared symbol tables and data profiles.
2. Method Validation: Screens candidates against a baseline and runs risk probes to check for data coverage and output degeneracy before full implementation.
3. Code and Review: Generates Python or MATLAB code and requires a JSON-based semantic review (syntax, reproducibility, etc.) to be saved to disk.
4. Results and Freeze: Produces a frozen_numbers.json file containing every number intended for the paper; any subsequent changes must be logged and re-frozen.
5. Paper Writing and Audits: Drafts sections based on the frozen snapshot and passes the paper through three independent auditors (consistency, completeness, and quality assurance) before submission.

Who it’s for

Students and researchers participating in mathematical modeling contests who want to maintain high rigor, avoid common clerical errors, and ensure their final submission is fully reproducible and compliant with contest rules.

Highlights

• Human-in-the-loop: Explicitly separates AI-owned mechanical correctness from user-owned modeling judgment.
• Frozen Numbers Convention: Prevents "stale" data by requiring all paper values to be traced to a frozen JSON snapshot.
• Multi-Auditor Layer: Requires three separate audits to pass before a paper is considered ready for submission.
• Native Plugin Support: Integrates directly as a plugin for Claude Code and Codex/ChatGPT.
• Risk Probing: Validates method feasibility and scale failures early in the process to avoid late-stage pivots.