Open leaderboard

Models on the fully-open question set — prompt and ground truth are public, so it’s reproducible but gameable. Live from the eval DB: every model is graded over the same bench, and a question a model didn’t validly answer is imputed with the worst Cramér any other model got on it (marked †).

Each board is an edition— a frozen snapshot of the question set. Every model shown answered that same set, so it’s a like-for-like campaign, not runs from different weeks stitched together. Adding questions doesn’t change a live edition; they roll into the next one, which a maintainer cuts deliberately. Older editions stay selectable as permanent snapshots.