Sakhanda Wire
NVDA $153.61 -1.62% MSFT $506.27 -2.26% GOOGL $280.41 -1.17% META $623.08 -0.53% AMZN $257.66 -0.03%
← Back to the news

ProvenanceGuard Verifies Whether MCP Agents Cite the Correct Source

Researchers introduced ProvenanceGuard, a post-generation verification layer for agents using the Model Context Protocol (MCP). Unlike conventional factuality checkers that pool all retrieved evidence, the system preserves each tool output’s source identity and checks whether a claim is supported by the source named or implied in the answer. This addresses cross-source conflation, in which a statement is true in the available evidence but attributed to the wrong record, document, or database.

The system decomposes answers into claims, identifies relevant sources, tests support with a natural-language inference model, checks numbers, dates and identifiers, and produces both claim-level verdicts and an overall allow-or-block decision. Blocked answers may be revised through a RARR-style repair loop and checked again.

In a medical-agent evaluation, the researchers analyzed 281 traces and had experts review 361 claims from 40 held-out answers. Experts marked 139 claims as unacceptable; ProvenanceGuard blocked 138 and missed one, while additionally sending 67 supported claims for review or repair. It identified the correct source for about 86% of claims with an identifiable source. In a test of 50 deliberately changed attributions, it detected all swaps. A harder evaluation involving similar sources produced an exact-source accuracy of 50.3%, highlighting a remaining limitation.

Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.

Originally published by Hugging Face on

Read the original on Hugging Face ↗

The reporting and the image belong to Hugging Face and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news