Authority Score

Authority Score is fundamentally a graph centrality metric, not a raw frequency count. It applies algorithms like PageRank variants or Eigenvector Centrality to the citation network. A case's score increases when it is cited by other high-authority cases, creating a recursive definition of influence. This distinguishes a landmark Supreme Court decision from a frequently cited but routinely distinguished lower-court outlier. Betweenness Centrality identifies cases that act as critical bridges between distinct doctrinal clusters, while closeness centrality measures how quickly a precedent's influence propagates through the entire legal network.

Not all citations are equal. Authority Score algorithms incorporate treatment type classification to weight citation edges. A citation labeled 'Followed' or 'Applied' transmits positive authority, strengthening the cited case's score. Conversely, a 'Negative Treatment' edge—such as 'Overruled', 'Criticized', or 'Questioned'—can diminish or even invert the authority flow. Distinguishing creates a neutral or weakly negative edge. This sentiment-aware propagation ensures the score reflects jurisprudential validity, not just visibility. A case that is frequently cited only to be criticized will have a lower score than one consistently relied upon.

Legal authority is not static. Authority Score models incorporate temporal citation analysis to account for precedent aging and influence velocity. A time-decay function can be applied to reduce the weight of older citations, reflecting that a case's current relevance may diminish if it is no longer actively engaged. Conversely, a citation cascade—a rapid, sustained increase in citations following a landmark decision—can be detected and used to boost the score. The model tracks citation velocity to distinguish a historically important but dormant case from a recently influential one that is actively shaping current litigation.

A raw global Authority Score is often legally meaningless. The computation must be constrained by jurisdictional filtering to reflect the hierarchical nature of precedent. A binding precedent from a higher court within the same jurisdiction receives a mandatory boost in the propagation algorithm, while a persuasive authority from an outside jurisdiction is weighted lower. The graph traversal is scoped to specific court hierarchies, ensuring that a case's score represents its influence within a particular sovereign legal system. This allows for the generation of jurisdiction-specific scores from the same underlying citation graph.

The final output is often a Precedent Influence Score, a composite metric that aggregates multiple signals. This includes the recursive graph centrality score, treatment-weighted sentiment, and temporal relevance. It may also factor in the seminal case detection signal, which identifies doctrinal origin points. The composite score provides a single, robust quantification of a case's total jurisprudential impact. This metric is used downstream for graph-based reranking in legal search, prioritizing not just semantically relevant documents, but those with high legal authority.

Advanced implementations use Graph Neural Networks (GNNs) to learn Authority Score as a node embedding. Instead of a manually designed algorithm, a GNN learns to generate a dense vector representation for each case that captures both its intrinsic features and its complex citation neighborhood structure. This approach can model non-linear authority relationships and incorporate heterogeneous node types, such as courts and judges, into a single heterogeneous graph. The resulting embeddings serve as a learned, high-dimensional Authority Score that can be fine-tuned for specific downstream tasks like case outcome prediction or link prediction.