Chain-of-Citation

The model does not merely cite a final case. It constructs a directed acyclic graph of authority, where each node is a legal source and each edge represents a logical dependency (e.g., 'relies on,' 'distinguishes,' or 'applies'). This prevents the common error of citing a case that does not actually support the proposition. The output is a verifiable path from a statute to a conclusion through intermediate precedent.

Before a citation enters the chain, the framework computationally verifies its precedential status. It maps the subsequent treatment history of a case to determine if its holdings have been overruled, questioned, or superseded. A chain is broken if a node relies on negative treatment, forcing the model to find valid alternative authority. This automates the manual Shepardizing process.

The framework uses an iterative search process where the answer to an initial query is used to formulate a secondary query to find connecting authority. For example:

• Hop 1: Retrieve the statute.
• Hop 2: Find cases interpreting the statute.
• Hop 3: Retrieve cases applying those interpretations to similar fact patterns. This enables the construction of a logical evidence chain that bridges a broad rule to a specific fact pattern.

A hard retrieval constraint limits search results to documents originating from a specific sovereign entity or geographic court system. The chain is jurisdictionally coherent; a federal circuit precedent is not used to incorrectly bind a state court unless proper vertical stare decisis applies. This prevents cross-jurisdictional contamination, a primary source of hallucination in general-purpose models.

A scoring function reduces the relevance of older legal documents to account for the evolution of statutory law and judicial interpretation. However, the weighting is non-linear: a foundational 1803 case like Marbury v. Madison retains high weight as binding precedent, while a 1950s interpretation of a frequently amended statute is decayed. This balances historical authority with legal currency.

Documents are not chunked by arbitrary token windows. A fine-grained chunking strategy segments legal documents into atomic, self-contained factual propositions. A single case might yield chunks like 'The court held that...' and 'The court distinguished...' This allows the citation chain to link to a specific holding within a case, not just the case as a whole, dramatically improving precise retrieval accuracy.