Skip to main content

BotBrain

BotBrain is an experimental neural controller behind ?botbrain=1. Without that flag, the game keeps using the scripted AI. The published model learns state-action pairs, but weights can only be replaced after passing the functional gate on seeds outside the bootstrap dataset and a manual review.

Testing the model​

npm run dev
# open http://localhost:4321/?botbrain=1
npm run bot:brain:check

In CAPTURE mode, the neural controller takes over combat when a target exists; with no target, the bot falls back to scripted navigation to capture and defend the points.

Collection and privacy​

Collection starts off. The player must opt in under Settings > Privacy > Help train the bots. In production:

  • UID + token authenticate the batch's origin;
  • the IP only takes part in rate limiting and is not stored in the corpus;
  • there are per-IP, per-player, and total-storage limits;
  • the importer caps each player's contribution;
  • no remote data publishes a model automatically.

Training locally​

npm i -D @tensorflow/tfjs-node
npm run bot:record 60 all
npm run bot:train -- --epochs=40
npm run bot:brain:check

The full operational guide, including Docker and the local sink, is in docs/BOTBRAIN-LOCAL.md. Docker exposes the game on loopback only; the local sink rejects external origins, caps rate, body size, and metadata, and stops collecting once it reaches 50 MiB.

Gates​

npm run eval:botbrain checks UID identity, consent, the CTF objective, cache busting, the local sink, non-root execution in the container, corpus balance, and separation between training and evaluation seeds. npm run bot:brain:check runs bot-versus-bot matches on holdout seeds and confirms the network moves, shoots, and gets kills.