TurnZero

The Illusion of Adaptation When Experts Lose — Stanford CS229, Winter 2025–26.

Players who just lost change their lead pair far more often than winners (72% vs 46%). This looks like adaptation, but it isn’t: 87% of post-loss switches don’t even interact with the opponent’s prior leads.

Using permutation-invariant transformer ensembles on 382K expert replays, a sequential model conditioned on prior-game outcome gains +6.6pp after wins but less than 1pp after losses. A specialist trained only on post-loss data performs worse than the base model on its own target subset. The post-loss signal actively misleads.

The live demo runs the full 5-member ensemble client-side via ONNX Runtime — no backend, no data leaves your browser.

See also: TurnOne — same dataset, different question.