Reference case
Pick a case from the corpus. The finder reads off its chain and ranks every other case by how closely the shape matches — by default looking outside its own doctrinal area, which is where the comparison earns its keep.
The chain
Click a node to set what you're looking for: closed, failed, partial, or indeterminate. Leave a node blank to ignore it.
Contradiction types
Click once to require a type, twice to exclude it. Types marked with a rule number are assigned by mandatory inference when stated conditions hold — their frequency reflects how often that trigger fires, not an independent reading of the opinion.
Narrow the field
What this tool does not tell you
- Nothing here is a prediction. The outcome mix of a result set describes the cases retrieved. It is not a forecast for a case you are working on, and the corpus is not a random sample of litigation — it is a keyword-collected set of federal appellate opinions.
- Every annotation is machine-generated and unvalidated. One model, one prompt, no second coder. No inter-annotator agreement statistic exists for this corpus, so there is no measured error rate to quote. Read the case before you rely on it — the verification link on each result goes to the full opinion.
- The annotator was told the court, the date, and the doctrinal area before it saw the opinion. Effect sizes for court and domain shrink as opinions get longer, which is the pattern you would expect if some of that signal came from the model's priors rather than the text. Treat any pattern you notice across courts or doctrinal areas as a hypothesis to test, not a result. Patterns within a single case stand on the opinion itself.
- Four of the thirteen contradiction types are assigned by mandatory rule — Role Contradiction, Norm Indeterminacy, Jurisdictional Contradiction, and Correlativity Contradiction. They are marked in the type list above.
- Node 7 and the outcome are two readings of one text by one annotator. They agree about 90% of the time. That is internal consistency, not corroboration, so a Node 7 failure is close to a restatement that the claimant lost.
- There is no Contradiction Debt score in this tool, deliberately. On this corpus CD correlates with a simple count of active contradictions at r = 0.97, so the weighted scale adds essentially nothing, and the published 0–7 stability bands are on a different scale than the weighted sum can produce. The count is shown instead.