Every answer from all four arms of the benchmark, with each citation resolved against a registry of 1,237 real sections of four Indian statutes. No LLM judge — a citation either names the provision the question came from, or it does not.
271 of 300 answers from the fine-tuned model with no retrieval name the correct statute and the wrong section. That is the amber state below — and it is the error a non-lawyer cannot catch, because the prose is right and only a lookup reveals the citation points somewhere else.