Three boards, one per visibility tier — a score on open, public, and private questions each mean different things. The credible capability ranking is the held-out (private) board; the others are reference and practice.
Models on the fully-open question set — prompt and ground truth both public. Reproducible by anyone, but gameable. A reference read, not a capability claim. Each board is a versioned edition: a frozen question set, so every model is compared on the same questions.
Open board →Community uploads scored against sealed ground truth. Provisional — the questions are visible, so treat it as a practice court.
Coming soonThe contamination-safe capability ranking — aggregate scores only, questions never released. Arrives once held-out questions exist.
Coming soon