A useful pattern for LLM classification at scale is to let it hallucinate plausible, fake entities. Then resolve to real ones.