Meta’s misstep reveals a deeper truth: evaluation-by-policy won’t protect us in production. AI safety won’t come from PDFs and policies, it has to be designed into the system.