StrictParity-GraphAudit: When Graph Retrieval Gains Do Not Become RAG Context Gains

In preparation. A graph-RAG evaluation paper on when graph retrieval gains deserve graph-specific credit under matched interfaces, leakage controls, and answer-boundary checks.

Authors: Iman YeckehZaare · Venue/status: EMNLP

This working manuscript introduces StrictParity-GraphAudit, a protocol for assigning graph-specific credit in retrieval-augmented generation under strict parity and leakage controls. On a multi-hop QA benchmark, graph retrieval improves exact-ID hit rate over flat dense retrieval, yet blinded human materiality and answer-generation checks do not favor the graph contexts and the gain inverts under train-edge holdout — licensing bounded exact-ID credit while blocking context-utility, generation, and full-system GraphRAG superiority claims.

Public artifact boundary: this route exposes status, authorship, visual summary, citation metadata, and the contribution boundary; manuscript files are posted only when review and prepublication rules allow it.

The rendered manuscript page adds status, visual summary, review boundary, citation metadata, and contribution notes. Key links: home, systems, papers, manuscripts, Google Scholar, GitHub, LinkedIn, ORCID, MIT profile, and CV PDF.