Ah yeah, the "hallucinations" classification is optimized for recall (keeping as many results as I can) not precision. It is mostly based on small models being the only support for a claim. Certainly lots of hallucinations everywhere!