We are classifying them at runtime, per entry. More accurate and LLMs are cheap.