Most retrieval-augmented systems optimize for a plausible-sounding answer. Legal AI can't afford that — a confident answer with no traceable source is worse than no answer at all, because it's the kind of mistake a practitioner won't catch until it's expensive.
What "citation-verified" actually means
It's not enough to show a source next to the answer. The system needs to verify that the cited passage actually supports the specific claim being made — not just that it's topically related. That check has to run before the answer is shown, not after, as a human review step that most users will skip.
Where this gets hard
- Hybrid retrieval (vector + keyword + graph) surfaces good candidates, but ranking them by evidential strength for a specific claim is a different problem than ranking by relevance.
- Legal language is precise on purpose — a citation that's "close enough" semantically can still be wrong in a way that matters.
- The verification step itself needs its own evaluation set, or you've just moved the trust problem one layer down.
Replace this with a real excerpt, finding, or lesson from your own work.
This is placeholder body copy. A real post here would probably walk through the actual verification pipeline, what failed in early iterations, and how the eval set for it was built.