Confidence Calibration. The ongoing comparison between claimed confidence and observed correctness, used to decide whether future confidence claims remain trustworthy.
Calibration keeps confidence from becoming decoration. A system that labels conclusions high, medium, or low confidence should later compare those claims with observed outcomes and review evidence. When high-confidence answers stop being reliably right, the system should retrain, investigate, lower trust, or tighten its confidence gate.
Calibration evidence can also feed an eval gate: a model that no longer calibrates honestly should not be promoted unchanged.