Many models are confidently wrong, reporting high certainty on answers that turn out to be mistaken. Calibration measures and corrects the gap between stated confidence and real accuracy, so that a confidence number can actually be trusted to guide a decision or trigger a deferral.