Model Kombat

A browser game where a simulated fly brain fights a language model in the 1990s arcade fighter mk.js. The README says the fight is instrumented well enough to say what each side contributes: real connectome wiring for the fly, and Jev answering typed choices about the same moves.

It is a 12,000-neuron sensory to descending to motor subgraph of the maleCNS v1.0 connectome with real wiring, synapse counts and transmitter signs, simulated spike by spike at 10 decisions per second. Named cell types like looming detectors LPLC2 and LC4 drive it, and the busiest motor pool picks the move.

Jev reads the fight as JSON and answers with one typed choice among the same seven moves, about five times a second. Nothing in the fly is trained in advance; learning happens during the fight through a dopamine-like rule, and Jev is one of several possible opponents you can select by URL parameter.

pip install -r requirements.txt, python scripts/fetch_game_assets.py to fetch the mk.js sprites, then python server.py and open http://127.0.0.1:5000/mk/. That runs everything that needs no account; to use Jev itself, put TYPESAFE_API_KEY in a .env file.

A softmax over 24 features trained on Jev's probabilities agrees with it on 93% of the calls it was confident about and answers in 4 µs. Run at that speed it beats the model it copied — 6.14 versus 5.69 damage per second — which the README highlights as a finding about latency rather than model quality.

The README calls reaction time the fly's whole edge: slowed to Jev's ~350 ms its damage output falls 70% and it never wins. Both convert decisions into moves at similar speed (an attack locks a fighter for 400–600 ms), but the fly can answer inside the 80–160 ms window when a strike is already in flight, and Jev cannot.

EXPERIMENTS.md holds thirteen experiments in the order they were run, failures and wrong predictions included. Standings and charts generated from 84 archived matches are on the project's GitHub Pages site; each learning setup and each source of decisions is entered as its own competitor, and every match writes telemetry JSON to telemetry/mk/.