Skip to content
DeicydePublic

About

Historical Lean autoformalization harness with human review, Verso source views, dependency graphs, and machine audits.

Topics

Resources

Stars

8 stars

Watchers

0 watching

Forks

Latest commit

 

History

302 Commits

Folders and files

Repository files navigation

Marathon v2

Human review for machine-generated mathematics.

License: MIT Python 3.11+ Historical release

Marathon explored how a mathematician could review and guide an automated Lean formalization. It connected Aristotle's prover to a browser-based review dashboard, local machine audits, and GitHub: inspect each declaration, check its dependencies, and send precise feedback into the next iteration.

It was developed alongside the GeometricAnalysis formalization of selected chapters of Lee's Introduction to Smooth Manifolds.

Try the demo · Architecture · Design history · Download v2.0.0

Historical release — November 6, 2026

This repository preserves the 2026 implementation as a research artifact. Aristotle no longer supports Lean-only projects, so the original automated proving and refinement loop cannot run against the current service. The review UI, existing review records, local Lean builds, and machine audits remain useful independently.

The historical SDK is pinned to aristotlelib==2.0.0. Upgrading the SDK alone does not restore Lean support. The setup guides describe the original workflow.

Marathon's declaration review dashboard, showing a synthetic Lean structure, its informal explanation, and review controls

Declaration review in the self-contained demo. A real project can supply a Verso build for semantic source rendering and definition lookup.

What's inside

  • Declaration review. Review definitions and theorem statements individually; keep, reject with a note, or defer. Decisions are tied to the code reviewed.
  • Context for the mathematics. Verso rendering, definition lookup, elaborated types, and whole-file or chapter views.
  • Dependency-aware navigation. Explore the project graph and review prerequisites before the results that use them.
  • Machine evidence. Local Lean elaboration, axiom and sorry accounting, statement dependencies, and checks for suspicious declaration patterns.
  • Durable orchestration. Isolated worktrees, resumable prover sessions, retries, publication recovery, and GitHub issue and pull-request workflows.

Statement review and proof completion were tracked separately: a human-approved statement could still contain sorry. The dashboard made that distinction visible.

Try the demo

Requires Python 3.11+ and uv.

git clone https://github.com/Deicyde/marathon.git
cd marathon
uv sync --locked
uv run --locked marathon deck --demo --no-open --port 8765

Open http://127.0.0.1:8765/. The demo uses synthetic review cards and simulated submissions; it needs no GitHub account, Aristotle key, or Lean project.

Use ← to reject, → to verify, space to defer, and ? for help. Opening an issue starts a declaration-by-declaration review; file and graph views provide the wider context in configured projects.

The original workflow

  1. Generate and audit. Produce Lean outlines, compile them locally, and collect evidence about their statements and dependencies.
  2. Review. Compare informal mathematics with the Lean declarations. Record local decisions and submit an issue's outcome explicitly.
  3. Refine and repeat. The conductor sends rejection notes to Aristotle, preserves work across iterations, and publishes changes through GitHub.

The conductor coordinated workers through project-local state, a SQLite ledger, Git worktrees, and GitHub. Human review remained a separate step from the prover.

Documentation

Read For
Architecture Module map, state ownership, and implementation invariants
Collaborator setup Configuring a project for the original review workflow
Shared deployment Hosting the dashboard behind GitHub authentication
Claude Code plugin The CLI commands exposed as a plugin
V2 design plan Design goals and development phases
Research and design reports Investigations, alternatives, and critiques retained as history

Use uv run --locked marathon --help to explore the preserved CLI.

Development

uv sync --locked --extra dev
uv run --locked --extra dev pytest -q

The v2.0.0 release passed 1,914 regression tests. The demo and tests are useful starting points for studying or adapting the harness without running a prover. Consumer-project state, credentials, and textbook sources are not distributed here.

License

MIT · Jack McCarthy, 2026.

About

Historical Lean autoformalization harness with human review, Verso source views, dependency graphs, and machine audits.

Topics

Resources

Stars

8 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages