Abstract
Legal doctrine does not change overnight. It drifts through a sequence of distinguished, doubted, and eventually overruled citations before a landmark judgment crystallises a new doctrinal position. Existing legal NLP treats doctrine as static, modelling individual cases in isolation and ignoring the temporal dynamics encoded in the citation graph. We present DRIFT, a framework that formalises doctrinal drift as a measurable signal on a temporal citation graph, introduces TGN-Law, a Temporal Graph Network adapted for legal citation semantics, and operationalises drift detection and inflection-point localisation as concrete evaluation tasks against a ground-truth corpus of 176 explicitly-overruled case pairs drawn from the RATIO Nigerian citation graph. TGN-Law employs treatment-type-aware message functions that assign distinct learned weight matrices to each of five edge types (cited, followed, applied, distinguished, overruled), combined with a GRU-based node memory that accumulates jurisprudential signals across decade-scale temporal snapshots. Drift is quantified as the cosine distance between consecutive concept centroid embeddings. Against a TF-IDF centroid drift baseline, TGN-Law achieves Precision@1 of 0.4489 versus 0.0170 (a 2,533% improvement) and NDCG@3 of 0.7502 versus 0.4945 (a 51.7% improvement) on drift-decade detection over 176 ground-truth inflection events. Attention analysis reveals that the model assigns disproportionately high attention to high-signal edge types (overruled, distinguished) relative to their frequency, confirming that TGN-Law recovers jurisprudentially meaningful structure automatically. We release all code, embeddings, and evaluation benchmarks under an open licence.