Confidence is estimated from token probabilities, entropy, cross-sample agreement, or the model's verbalized confidence.
Models often answer confidently even when wrong (poor calibration, hallucinations).
