Math To Manim

Every Dial Has a Loose Switch

A 10-minute film about the proof of the Aaronson–Ambainis conjecture, made by the Mythos chain in this repo. The conjecture asked whether a quantum computer can be exponentially faster than an ordinary one on random inputs. The proof shows it cannot: every bounded polynomial of degree d has a bit whose influence is at least c·v^{5/2}/d^6, with c = 8√15/(125π⁴) ≈ 0.0025. (announcement and Lean proof)

How it was made:

• The input was a written film brief: the intent, the prerequisites, the exact formulas and a worked example with checked numbers, a palette, and nine scenes.
• The six Mythos stages (intent, cartographer, curriculum, math director, cinematographer, scene composer) and codegen ran on Claude Opus 5.5.
• The chain wrote one Manim ThreeDScene, LooseSwitchJourney (964 lines). One repair pass restored the paper background that the brief asked for.
• Manim rendered 586 animations into the film. The GIF above is 31 seconds of it.

Twelve frames, one every 50 seconds:

Fable 5.2 Learns Manim

Claude Fable 5.2 wrote this week's changes to the Mythos chain, and the images in this section. The title image is a still from a Manim ThreeDScene that the model wrote: docs/assets/fable-5-2/src/hero.py. The chain itself calls the model set in M2M_MODEL (default claude-fable-5-2).

What changed in the chain

Two commits changed how a run reports its state.

a49352e: explicit job and run lifecycles. A job's status now moves only through a transition table: queued to running, then completed or failed. The run manifest records three fields under status:

• run: running, completed or failed.
• stage: the step that is running now. On a failed run, the step that failed.
• outcome: how a finished run ended. One of film_ready, scene_ready, validation_failed, render_failed, render_timed_out, render_blocked, chain_failed.

Before this commit, a crashed run looked the same on disk as a run in progress. Now it is recorded as failed, with the step that failed.

9e154fd: serve-mcp runs the Mythos chain and returns the Mythos job. src/mythos/mcp_server.py is now an MCP server over the Mythos service. It has seven tools: m2m_create_animation, m2m_get_job, m2m_list_runs, m2m_get_run, m2m_get_artifact, m2m_get_scene_code, m2m_cinematic_charter. An agent that polls a job gets status, stage and outcome in the job record. A completed job can still have the outcome render_failed, so read outcome to know if a film exists. The Grok chain has its own server: math-to-manim-grok serve-mcp. No silo imports another silo (tests/test_silo_boundaries.py).

The job records in both diagrams are real output. capture_job_records.py calls the MCP tool functions with a stub model backend that stops at the fourth stage.

How it works now

The chain has six reasoning stages. Each stage writes one JSON file into the run directory. Then codegen writes the Manim scene, verify checks it, and render runs Manim. If a check or a render fails, repair asks the model to rewrite the scene file. The default limit is 3 repairs.

Run the chain without a model, to test the setup:

math-to-manim-mythos run "Why does a top precess?" --offline

Run the chain with the model and render the film:

This is the beginning of the README. Read the rest at the original. Read the full README ↗