Abstract
We present RATIO (Retrieval of Authority via Typed-edge Inference and Ontology), a graph-aware retrieval framework for legal precedent search that explicitly models the citation network rather than relying solely on semantic similarity. Current legal retrieval-augmented generation (RAG) systems retrieve document chunks by embedding proximity, which cannot distinguish binding precedent from overruled authority, foreign persuasive authority, or cases cited for propositions they actually reject. We formalise the Binding Horizon Problem, the observation that valid precedent retrieval is a temporally-constrained, jurisdiction-masked graph reachability problem, not a static nearest-neighbour search. We construct the first open citation graph for Nigerian case law, spanning 14,437 Supreme Court and Court of Appeal judgments with 51,465 resolved citation edges typed as applied, distinguished, overruled, followed, or cited, extracted from 317,476 raw citation mentions across ten citation formats. We implement the binding horizon filter as a live Stage 2 component of the retrieval pipeline, operationalising temporal validity and jurisdictional binding as executable graph masks. We train a Graph Attention Network (GAT) that fuses 387-dimensional semantic and structural node features with typed edge attributes, achieving a +184.0% improvement in Hits@10 and +271.3% improvement in MRR over a non-graph MLP baseline on held-out citation prediction. We release the citation graph, trained embeddings, and evaluation benchmark under open licence.