Probabilistic predictors that answer queries one at a time face a practical problem: the ground-truth labels needed to check calibration are often missing. This paper introduces a way to certify calibration without such labels, by exploiting the logical coherence among the questions the predictor is asked.

The authors frame the problem in terms of a bettor whose stakes may be restricted in various ways. For each restriction, they prove a minimax theorem that characterises when calibration can be certified. The resulting certificates are described as sharp and stable, meaning they are tight and do not depend on fragile assumptions.

Because the approach is label-free, it could apply to settings where obtaining ground truth is expensive or impossible, as long as the questions are logically related. The contribution is theoretical, establishing the conditions under which such certificates exist.