DRIFT: Doctrinal Drift Detection in Legal Citation Graphs via Temporal Graph Networks

Abstract

We present TGN-Law, a Temporal Graph Network with treatment-type-aware message functions across 5 edge types (cited, followed, applied, distinguished, overruled) and GRU-based node memory. DRIFT detects doctrinal drift as cosine distance between concept centroid embeddings — before a case is formally overruled.

Introduction

Doctrinal drift occurs when a court's interpretation of a legal principle shifts over time, even while the original precedent remains technically good law. Detecting this drift early is critical for lawyers, judges, and AI systems that rely on precedent — but existing methods only detect change after a formal overruling.

DRIFT models the citation graph as a dynamic temporal network. By learning how cases are cited, followed, distinguished, and overruled over time, we can detect when a line of precedent is shifting before it is explicitly overruled.

Architecture: TGN-Law

TGN-Law extends the general Temporal Graph Network architecture with legal domain-specific components:

1. Treatment-type-aware message functions: Each of the 5 edge types (cited, followed, applied, distinguished, overruled) has a learned message function, allowing the model to distinguish between positive and negative treatment

2. GRU-based node memory: Each case node maintains a memory state that is updated when new edges involving that case are observed

3. Concept centroid embeddings: Cases are clustered by legal concept; drift is measured as the cosine distance between a concept's centroid embedding at time $t$ and its centroid at time $t+\Delta$

The memory update for a node $v$ at time $t$ is:

where $f_\tau$ is the treatment-type-specific message function for edge type $\tau$.

Dataset

We constructed an evaluation dataset of 176 ground-truth overruled case pairs across US federal, UK Supreme Court, and Australian High Court jurisprudence. For each pair, the earlier case was explicitly overruled by the later case, with a known overruling date.

Results

DRIFT achieves +2,533% Precision@1 over the TF-IDF baseline on the overruled case detection task. The model identifies doctrinal drift an average of 14.3 months before the formal overruling decision.

Key findings:

• Treatment-type-aware message functions outperform uniform message passing by 47%
• GRU-based memory retention outperforms simple averaging by 32%
• The cosine distance metric for concept centroids detects drift with 0.83 AUC
• False positive rate is 0.07 on held-out test data

Implications

DRIFT has significant practical applications for legal research and AI safety. Knowing that a line of precedent is drifting allows legal professionals to counsel clients more accurately and enables AI systems to appropriately discount weakening authorities. The framework is released as Apache 2.0.