Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents

Hugging Face has introduced ProvenanceGuard, a source-aware factuality verification layer designed for Model Context Protocol (MCP) agents. Traditional verification systems check whether a claim is supported by pooled evidence, but they typically fail to verify which specific tool output supports each claim or whether the answer attributes the claim to the correct source. To address this gap, ProvenanceGuard targets cross-source conflation, a failure mode where a claim is true somewhere in the evidence pool but incorrectly attributed. Operating as a post-generation layer on top of black-box MCP agents, ProvenanceGuard preserves source identity through a pipeline that breaks down answers into specific claims, finds relevant sources, checks support, compares sources against named attributions, and emits verdicts or allow and block decisions. In testing on 281 real traces from a medical agent using patient records and research articles, human experts identified 139 claims that should not pass, and ProvenanceGuard caught 138 of them. It also held 67 supported claims for review, reflecting a conservative design policy tailored for data-sensitive environments. When paired with a RARR-style repair loop, blocked answers were resolved through source-grounded revisions or safe fallback text. While the system effectively caught misattributions, tests involving multiple similar sources highlighted that distinguishing between closely related sources remains a challenge. The verification process introduces a modest overhead of roughly half a second per answer in the evaluated local setup, which utilized models such as MiniLM and DeBERTa. According to the excerpt, the approach can be adapted to other fields and hosted models wherever agent traces preserve tool outputs and source IDs.