Quantifying model confidence or epistemic/aleatoric uncertainty in predictions; used for calibration, abstention, and risk‑aware decisions.