sweetcornna/mathodology

What it solves

Mathematical modeling contests require a structured workflow: reading the problem correctly, building explainable models, testing conclusions with experiments, producing clear figures, citing only real references, and writing a readable paper. Mathodology packages that workflow into reusable skills, roles, templates, and a desktop app so agentic coding tools like Claude Code, Codex, and Antigravity CLI can do it consistently.

How it works

Mathodology provides eight skills (e.g., modeling prompts, evidence retrieval, scientific plotting, review questions) and eight professional roles (coordinator, problem analyzer, modeler, programmer, literature, reviewer, writer, packager) that run in parallel on demand. It includes 46 chart templates written in Matplotlib — fill in real data to get PNG/PDF output — plus 18 experiment components for common math-modeling procedures like sensitivity analysis, Monte Carlo, backtesting, ablation, and ranking stability. The verify_refs.py script checks each citation for DOI, title, author, year, and retraction status. The Mathodology Studio desktop app (macOS/Windows/Linux) drives locally installed agent CLIs without calling model APIs, storing every chart/table/code change as a version for comparison and traceability. It also provides a read-only review mode that maps conclusions to tool logs, a literature panel that writes references.bib after DOI verification, and an optional humanizer to reduce AI-typical phrasing.

Who it's for

Competitors and teams in mathematical modeling contests who want to use AI coding agents but need a structured, reproducible, and citation-safe approach. Also suits anyone who wants their coding agent to follow a rigorous modeling pipeline — from understanding the prompt to delivering charts and a clean paper.

Highlights

• 8 skills and 8 roles cover the full modeling workflow, from problem reading to packaging.
• 46 chart templates with data requirements, layouts, misleading examples, and caption prompts.
• 18 experiment components integrate with charts and produce three-line tables.
• verify_refs.py checks each reference for DOI, title, author, year, and retraction — flags suspicious IDs.
• Studio desktop app version-controls every chart, table, and code change, and traces them back to the agent turn that produced them.
• Two-tab citation review: trace each conclusion to tool logs, and verify every reference one by one.
• Works with many agent CLIs (Claude Code, Codex, Antigravity CLI, Qoder CLI, pi, etc.) via their own interfaces, switchable between turns.
• Open source under MIT; no model API call, no key storage, no data collection in the Studio app.