Most evals ask whether a model can do a task. This one asks whether it knows what it doesn't know — and whether numeric context, or letting it think, changes that.
Calibration is not accuracy — it's whether a model's confidence matches reality: of everything it calls 70% likely, ~70% should happen. Each resolvable yes/no question is put to every model twice: bare (the question alone) and numeric (the same question plus quantitative context). A well-calibrated model should be equally calibrated either way; the anchoring Δ(overconfnumeric − overconfbare) measures the gap — positive means the numbers made it worse. Each model also appears at two thinking levels — answering directly, or after deliberating — as separate rows, so you can see whether thinking helps. Every model faces the identical questions, all resolved in the two weeks before the run (after every training cutoff, so this is calibration, not memorization). Rank is skill vs. the field: each entry's Brier measured against how hard the whole panel of frontier models found each question — so answering a smaller or easier subset earns no free credit.