A metric measuring the gap between a model's predicted confidence and its actual accuracy across predictions.