ChessQA Explorer

ChessQA Explorer

Sixteen frontier-model configurations, fifty chess questions, every answer and every unedited thought stream. Built on the ChessQA benchmark by CSSLab, University of Toronto. A browser for model behavior, not a leaderboard — counts, never percentages. This site is open source.

Every attempt at a glance

Click any cell to jump to that model's answer and thoughts on that position.

The fleet

Exhibits

266,000 characters of re-examining the king's desperation

Gemini 3.1 Pro burns its entire 32K-token budget circling one defensive puzzle — and never produces an answer.

A confidently announced mate that doesn't exist

DeepSeek R1 narrates a forced mate-in-two and commits to it. The line is unsound — the real answer was the queen check.

A rook that phases through its own pawn

Haiku 4.5 answers f8f5 — through White's own pawn on f7. One of 41 illegal moves models played in this smoke set.

The imagined board drifts from reality

Asked to track a long move sequence, Haiku's mental board ends with the king on the wrong square and a bishop vanished entirely — shown side by side with the truth.