Stare Decisis Modeling

The foundational distinction in precedent modeling. Binding precedent originates from a higher court within the same jurisdictional hierarchy and imposes a mandatory constraint on the lower court. Persuasive authority comes from courts outside the binding hierarchy—such as sister circuits or lower courts—and carries only advisory weight. Computational models encode this as a binary or weighted edge attribute in the citation graph, with jurisdictional filtering algorithms constraining traversal to only legally relevant nodes.

A structured representation of court levels that determines the direction and force of precedential obligation. The model maps the vertical hierarchy—from trial courts through appellate courts to a supreme tribunal—and the horizontal geography of circuits and districts. Key elements include:

• Vertical binding: Lower courts must follow higher courts within the same chain
• Horizontal non-binding: District courts are not bound by other district courts
• Circuit boundaries: Federal appellate decisions bind only district courts within their geographic circuit This hierarchy is encoded as a directed acyclic graph (DAG) that constrains authority propagation algorithms.

A quantitative scoring mechanism that computes the precedential force of a prior decision for a given current case. The calculation aggregates multiple signals:

• Jurisdictional proximity: The hierarchical distance between the citing and cited courts
• Factual similarity: Semantic overlap between case fact patterns, measured via embedding cosine similarity
• Treatment history: Whether the precedent has been subsequently followed, distinguished, or overruled
• Temporal decay: A time-weighted factor that may reduce the weight of very old decisions unless they are landmark cases The resulting score informs downstream tasks like case outcome prediction and citation recommendation.

The computational modeling of a court's decision to not apply a seemingly binding precedent by identifying material differences in facts or legal issues. This is a critical edge case in stare decisis modeling because it represents a lawful departure from obligation. Models capture this through:

• Fact vector comparison: Embedding the facts of the precedent and the current case, then measuring divergence
• Materiality classification: An NLP classifier trained to determine whether identified factual differences are legally material or merely superficial
• Distinguishing edge: A specific edge type in the citation graph indicating the precedent was cited but not followed, preserving the network's integrity while accurately reflecting judicial behavior

The modeling of how precedential authority evolves over time. Legal doctrines are not static; they age, gain influence, or are abruptly terminated by overruling events. Key temporal mechanisms include:

• Citation velocity: The rate at which a decision accumulates citations over time, used to detect rising or waning influence
• Precedent aging curves: Statistical models of how citation frequency typically decays, with landmark cases exhibiting anomalous persistence
• Overruling event detection: Real-time monitoring of the citation graph for negative treatment signals that terminate a precedent's authority
• Doctrinal drift tracking: Measuring how a legal principle's interpretation shifts across a chain of applying decisions using semantic drift detection on opinion text

Graph algorithms adapted from network science to distribute precedential influence scores across the citation graph. Common approaches include:

• PageRank variants: Iterative algorithms where a node's authority score is the weighted sum of scores from citing nodes, with jurisdictional constraints on edge traversal
• HITS (Hyperlink-Induced Topic Search): Separates nodes into hubs (cases that cite many important authorities) and authorities (cases cited by many important hubs)
• Personalized PageRank: Biases the random walk restart probability toward nodes in the relevant jurisdiction, ensuring scores reflect binding rather than general influence These algorithms produce the Authority Score that quantifies each case's precedential weight.

A directed network structure where nodes represent legal cases or statutes and edges represent citation relationships. This forms the foundational data structure for computational precedent analysis.

• Nodes can be enriched with metadata: court level, jurisdiction, date, and full text embeddings.
• Edge attributes include treatment type (e.g., 'overruled', 'followed') and citation sentiment.
• Enables graph algorithms like PageRank variants to compute authority scores.

The process of using a citator tool to trace a legal authority's subsequent treatment history. The goal is to determine whether a case remains 'good law'—meaning it has not been overruled, criticized, or otherwise negatively treated.

• Originates from Shepard's Citations, now a standard feature in Westlaw and LexisNexis.
• Computational shepardizing automates this by traversing the citation graph and classifying treatment types.
• Critical for maintaining the integrity of any AI-driven legal reasoning system.

A graph algorithm that iteratively distributes precedential influence scores across a citation network. It identifies the most legally significant nodes by analyzing both the quantity and quality of citations.

• Often uses PageRank variants adapted for legal hierarchies (e.g., weighting citations from higher courts more heavily).
• Incorporates jurisdictional filtering to ensure scores reflect only legally relevant precedent.
• The output is a ranked list of cases by their computed precedential weight.

An NLP task that automatically categorizes how a citing case legally treats a cited authority. Labels include 'overruled,' 'distinguished,' 'followed,' or 'criticized.'

• Relies on fine-tuned language models trained on annotated citation sentences.
• A critical input for negative treatment detection and authority score decay.
• Enables automated flagging of cases that have been weakened or invalidated.

A fundamental distinction in legal reasoning that computational models must encode.

• Binding Precedent: A decision from a higher court within the same jurisdiction that a lower court is legally obligated to follow. Modeled as a mandatory constraint with a specific jurisdictional scope.
• Persuasive Authority: A decision from outside the binding hierarchy that a judge may consider but is not required to follow. Weighted lower in authority propagation algorithms.
• Jurisdictional filtering enforces these constraints during graph traversal.

A deep learning architecture designed to operate directly on graph-structured data. In legal AI, GNNs learn node embeddings that capture both a case's intrinsic features and its citation neighborhood structure.

• Used for tasks like link prediction (forecasting future citations) and node classification (identifying seminal cases).
• Operates on heterogeneous graphs that model cases, statutes, courts, and judges as distinct node types.
• Enables more nuanced authority scoring than traditional PageRank by learning complex, non-linear propagation patterns.