OpenMath — Catch-Up Agent

An agent that turns an AI-generated math paper into a playable picture, unattended. Built at the "Build an Agent" Hackathon, Corgi Cafe, San Francisco, Oct 7, 2026.

Status: working on one paper (Family 025). One command runs the audit on Agent37, scores it, searches prior work through Monid, and publishes the card to Supabase.

Demo video: https://drive.google.com/file/d/1Heyey2PuSMREzx5KD7WMdONOz1z_djNZ/view?usp=sharing Live app, all features: https://reliance-casino-judicial-domains.trycloudflare.com (a tunnel to the demo machine; up while that machine is on) Static pages: https://elmdin.github.io/OpenMath/ (uses the same server for chat and experiments)

What it does (target)

Paper in -> briefing out, no human step:

1. Claim in plain words.
2. Verification status: Lean-checked / human-reviewed / unreviewed (read from the source repo, not judged by us).
3. An interactive picture: the agent chooses a simulation from a small vetted library and emits a JSON spec. It never emits pixels or code.
4. A self-check: exact arithmetic on the paper's worked example.
5. Open questions.

Stack

• Agent37 Cloud APIs: the unattended worker (required by the hackathon)
• OpenAI: structured output for the briefing and the picture spec
• Supabase: stores finished briefings; the web page reads the feed

Layout

• data/family-025/ first target paper (OpenAI, "Short Egyptian fractions"), see SOURCE.md
• docs/SETUP.md accounts, keys and the pre-build checklist
• docs/BRIEF.md the first paper, the picture, and the demo script
• docs/DECISIONS.md what we decided and why; links to the full research doc
• docs/REQUIREMENTS.md hackathon rules, sponsor plan, submission checklist
• supabase/schema.sql
• docs/RESEARCH.md what we found online, what Family 025 looks like, ideas kept and dropped
• agent/ proof map, exact check and card builder (python3 -m pytest agent -q)
• web/ the card page; build data with python -m agent.card data/family-025 web/card.js

Run it

cp .env.example .env            # fill in the keys (MONID_API_KEY for the prior-work search)
python3 -m pytest agent -q      # 70 tests
python3 -m agent.run data/family-025 "Family 025: Short Egyptian fractions" agent37   # or: openai
python3 -m agent.server         # then open http://127.0.0.1:8765 (summary card, review copilot, chat)

See docs/SUBMISSION.md for the pitch and docs/RESEARCH.md for the research behind it.