Authority-Aware GraphRAG for Explainable Tax-Risk Evidence Retrieval: A Controlled Evaluation of Typed Legal Graph Expansion

Tax-risk analysis needs evidence sets that include a governing provision, companion rules, and a visible relationship path, rather than a single high-scoring passage. We study the retrieval stage of graph-grounded tax analysis through an executable, controlled benchmark. Building on citation-graph expansion used in tax-oriented retrieval systems, we propose Typed Evidence-Budgeted GraphRAG (TEB-GraphRAG), which jointly scores lexical seeds, relation types, and a deterministic sparse two-topic gate while preserving graph-path provenance. A transparent synthetic collection contains 59 real tax-section identifiers with researcher-authored synopses, 238 researcher-authored directed relations, and 700 templated issue descriptions; 500 held-out queries are scored without learned retriever training. On this benchmark, our method reaches 85.40% Recall@5, versus 69.27% for unrestricted one-hop graph expansion and 59.93% for BM25. The 16.13-point advantage over graph expansion has a topic-cluster bootstrap 95% interval of [11.37, 21.60] points. Under type-label permutation on a balanced 200-case stress subset, mean Recall@5 drops to 0.597; an unrelated-topic distractor reduces full-test recall to 0.636. Typed retrieval therefore depends strongly on metadata fidelity and query focus. These controlled findings do not establish legal correctness, tax-risk classification, or LLM answer quality.