Transitive Authority

Testing whether models can reason across three-case citation chains—a capability no other legal benchmark measures.

Case A

ANCHOR

Original precedent

Case B

MIDDLE

Cites A

Case C

NEWEST

Cites B → predicts A

The Challenge

Given Case A (anchor) cited by Case B (middle) cited by Case C (newest), can the model predict how C treats A—without being told directly? This requires chaining inferences: B→A treatment plus C→B treatment should inform C→A treatment.

[ REASONING CHAIN ]

If B follows A...

And C follows B...

Then C likely follows A

Budget Distribution

Opinion text is distributed strategically: 40% for the newest case (C), 60% split between anchor and middle. This forces the model to reason with incomplete information—mirroring how lawyers work with case summaries before reading full opinions.

Novel Evaluation

No other legal AI benchmark tests transitive reasoning across precedent chains. This evaluates a fundamental capability lawyers use daily: inferring relationships between authorities through intermediate citations.