Build a graph from verified records, and compare with an LLM-extracted one

Both graphs use one schema (matter, title, person, company and clause nodes), so the same query functions answer the 15 questions on each and any difference comes from how the graph was built. The decision that matters most is that entity resolution is a separate, swappable Resolver. Graph B is built twice from the same extractions: once keyed on names exactly as printed, and once with rules (company numbers, aliases learned from other registers, normalised names). The gap between those two rows is the answer to "what rules would close the gap", measured rather than argued.