The method keeps tool-output source IDs separate rather than pooling evidence, then decomposes an answer into claims, routes each claim to a source, checks support and attribution, and returns per-claim verdicts plus an allow-or-block decision.
The reported evaluation used 281 traces from a medical agent. Human experts reviewed 361 claims from 40 held-out answers: 139 claims were judged unsuitable to pass, and ProvenanceGuard caught 138; it also held 67 claims the experts judged supported. For claims with an identifiable source, the article reports about 86% exact-source selection. On a 50-case controlled attribution-swap test, it detected all swaps.
The local experimental configuration used MiniLM for source routing, a DeBERTa NLI model for support checks, and a local language model for claim decomposition. The article reports roughly half a second of overhead per answer. It does not establish performance on production deployments, hosted-model configurations, or domains beyond the reported medical-agent tests; it also reports 50.3% exact-source identification in a harder similar-source test.