Recording hard constraints on AI use that favorable test results cannot override, prioritizing safety limits over performance gains.