Fine-tuning taught the model which statute governs a question (47.7% → 92.1%) and almost nothing about which section (0.3% → 0.9%). It names a real section 99% of the time and the correct one about 1% of the time.
What is live here: the citation checker below runs the benchmark's own regex and section registry in your browser — paste anything and it resolves against 1,237 real sections. No model, no server.
What is not: the 7B generation. A free Space cannot serve a 4-bit 7B, and quietly swapping in a smaller model would demo something this project never measured. So this serves the real measured answers — 1,200 responses from the actual run — and the real retrieval the RAG arms were given, computed with the same bge-large embedder.