Citation graph
BetaForward and backward citation chains, with a treatment label on every edge.
- ✓Forward and backward chains, plus a depth-limited graph walk and multi-seed exploration
- ✓Treatment label on every edge, and a most-overruled listing
- ✓Filter a graph query by treatment
- ✓Export the graph data as CSV or JSON
What it is.
Every judgment is a node; every citation is an edge carrying a treatment — followed, distinguished, overruled, referred, or relied on. You can walk backwards to what a case cited, forwards to what cited it, run a breadth-first expansion to depth three, or explore from several seed cases at once.
Read the coverage caveat first. An edge can only be created when we hold the citing judgment's text and a matching parallel citation on the cited judgment, and both of those come from the full-text ingest route that is not switched on in production. The machinery is real and the graph is close to empty. Nothing in the product pretends otherwise: when the citation table holds no data, a judgment card says in terms that subsequent treatment was not checked, rather than implying the case is good law.
Three steps.
End to end.
Open any judgment in the index.
Backward to what it cited, forward to what cited it, or a depth-limited graph walk from one or several seeds.
Each edge is labelled followed, distinguished, overruled, referred or relied on. Filter a graph walk to the treatments you care about.
What you get.
- ✓Forward and backward chains, plus a depth-limited graph walk and multi-seed exploration
- ✓Treatment label on every edge, and a most-overruled listing
- ✓Filter a graph query by treatment
- ✓Export the graph data as CSV or JSON
- ✓Explicit "treatment not checked" wording whenever the citation graph holds no data, rather than silence that reads as reassurance
Quick answers.
Two automated passes and no human review. On insert, a keyword heuristic reads a window around the citation and defaults to "referred" — crude enough that a sentence declining to overrule can be labelled overruled. On first view, a Claude pass re-classifies edges in batches and records a confidence. There is no editorial review step, and we should not claim one.
No — CSV and JSON only. Rendering is done in the app.
No. The only graph filter is treatment. There is no doctrine or subject filter on the citation network.