AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries