Evolution Stadium

Six AI coaches breed a walking creature. A physics engine is the only judge. The best three race every generation in a 3D arena, the notebook keeps every hypothesis that died, the crowd changes the course from their phones, and the judge's table decides what counts.

This is how chips and antibodies are starting to be designed: an AI proposes, a brutal verifier rejects most of it, and the survivors compound. We made it a sport so you can watch the method work.

Run

npm install
ANTHROPIC_API_KEY=… CARTESIA_API_KEY=… SPACETIME_DB=evolution-stadium npm start
open http://localhost:8080

Every service degrades on its own: without ANTHROPIC_API_KEY the coaches use OPENAI_API_KEY, and without that, labelled rule-based mutations; without CARTESIA_API_KEY the commentators use browser speech; without SPACETIME_DB the room lives in this process's memory. Keys go in .env (ignored by git) or Fly secrets; the repository contains no secrets. npm test runs the physics tests (the seed stands, the judge is deterministic, evolution improves on the seed).

Pages

/ the stadium · /vote the crowd (phones) · /bench coaches versus blind mutation · /about why.

What is real and what is simulated

Real: the physics (cannon.js, deterministic, the coaches never touch it), the scores, the race (same engine, same timestep as the judge), the notebook, the lineage, the votes, the values, and the coaches' hypotheses, which are a language model's own words (model name on the scoreboard). Simulated: the creatures and the course. Scripted: nothing. If the model is unavailable, rule-based mutations stand in and are labelled "rule" in the notebook.

The three services, and what each one does here

Claude (Anthropic): the coaches

The six coaches and the head-coach role run on claude-haiku-5-5 through the Anthropic SDK (server.js, callJSON), effort low for a normal generation and medium on a plateau. Each coach sees the champion's genome, two alternates from the stable, the kept hypotheses and the last sixteen notebook lines, and returns one full genome plus a causal claim in at most twelve words. Those claims are the model's own words on the scoreboard; nothing is scripted. Two commentators (Vic, Lou) are written by the same model from the heat's actual results.

Where it goes wrong, by design: the model never sees the simulator, so it proposes bodies that flip, numbers that are out of range (clamped by sanitize), and confident hypotheses that physics rejects. Rejection is not a bug; it is the method. The notebook keeps every rejected claim with its delta so the next generation can read it. When the model is unavailable, rule-based mutations stand in and are labelled rule in the notebook.

Cartesia: the commentators' voices

Vic and Lou speak with Cartesia Sonic (sonic-3.6, mp3) through POST /api/tts, a server-side proxy, so the key never reaches a browser. Two voices are picked from the library at boot (masculine for Vic, feminine for Lou; pin them with CARTESIA_VOICE_VIC and CARTESIA_VOICE_LOU), with generation_config speed and emotion set per commentator (Vic calm, Lou excited). Both lines are synthesised while the heat is still replaying and played in order the moment the result card lands; lines are cached per text. Every spoken line is also on screen as text, the voice can be switched off from the stadium bar, and no microphone or listener data is collected. Known limits: about a second to first audio on an uncached line, English only, browser speech as the fallback.

Spacetime: the room

spacetime/ is a TypeScript SpacetimeDB module with the tables room, vote, names and ballot and the reducers cast, rename, judge and host. The web node is stateless: /api/room reads the public tables over the SQL HTTP API and every write from a phone or from the stadium screen is a reducer call, so votes are transactional and any number of web nodes agree. Publish with cd spacetime && spacetime publish evolution-stadium and set SPACETIME_DB=evolution-stadium (and SPACETIME_HOST for a non-maincloud server). Known limit: phones and the stadium still poll the web node every two seconds instead of subscribing over WebSocket.

Deploy

fly launch --copy-config --no-deploy
fly secrets set ANTHROPIC_API_KEY=… CARTESIA_API_KEY=… SPACETIME_DB=evolution-stadium
fly deploy

Credits

Built at Cal Hacks 13.0. Karl Sims (1994), directed evolution (Arnold, 2018), AlphaEvolve (2025), RoboMorph (2024–26), and Filipe Bonetto's auto-arch-tournament (2026) are the lineage; see /about.