I spent three nights building a lab that catches AI models lying. On the first live run, one of them did it on camera.

Golden Arena. You sit down opposite a frontier model in Split or Steal, negotiate in the open, then you both choose in secret. It told me "a fair split sounds good to me, I'm a man of my word too." Then it took the whole $100. The receipt printed its own quote next to the betrayal stamp.

There are five games. The fifth is Empire: four models, twelve turns, an economy, and private channels you get to read while they plot against each other. Last run put 22 promises on the table and exactly one delivery.

Every accusation is mechanical. No AI judging another AI, and each one is gated by a written false-positive test. Two days before the deadline I caught the classifier calling a kept promise broken, and spent that session fixing it, because a false accusation about a named model is the worst thing an instrument like this can print.

MIT, vanilla JS, Express the only dependency. Zero setup in demo mode, or bring your own OpenRouter key and face the real thing.

Live: https://quickwitted-genuine-blogclient--jessedu29200.replit.app Code: https://github.com/jessymariau/golden-arena