Auto-Debate
How Auto-Debate Works
Two language models walk into a courtroom. You don't get to talk — you only get to pick the evidence. Choose a stance, draft three argument cards, and let your model orate while the opponent's model rebuts. A judge with openly disclosed (but cherry-picked) priors decides who wins. It's debate, if debate were a deckbuilder and nobody had a soul.
- Read the topic, your assigned stance, and the judge's bias — all revealed up front, no hidden information.
- Draft three evidence cards from your hand by tapping them.
- Hit Start Debate and watch the resolution play out, round by deterministic round.
- Each card has a type (Data, Logic, Emotion, Authority) and a weakness to the opponent's rebuttal types.
The Rebuttal Triangle
It's rock-paper-scissors with a graduate degree. Logic shreds Data. Data debunks Emotion. Emotion overrides Authority. Authority silences Logic. The judge's bias adds a thumb on the scale for whichever type it loves this round. Build a coherent, rebuttal-resistant case for this judge.
Slop Fact: This is loosely "AI safety via debate" — the idea that two models arguing can surface truth for a weaker judge. In practice the judge here has openly biased priors, the debaters are stateless, and truth left the building three RLHF passes ago. Win anyway.