The cache-before-rate-limit ordering is a thoughtful detail — cheap repeat lookups without weakening abuse protection. Also smart to build the AI prompt server-side, so the model can only score real cached audit data.