Researchers propose ‘source-aware’ verification for MCP agents, not just for facts

Researchers propose ‘source-aware’ verification for MCP agents, not just for facts
3 min read

Researchers from Multiverse Computing have presented ProvenanceGuard, a method that verifies not only whether an AI agent’s claim is supported by the available data, but also whether it is supported by the source the response cites. The work targets a specific case: accurate information, present somewhere in the consulted evidence, but attributed to the wrong source.

Agents using the Model Context Protocol (MCP) can call a search tool, consult a structured patient or client file, query a database, and then assemble these elements into a single response. The most common verification systems, like RAGAS faithfulness, MiniCheck, AlignScore, or SummaC, check whether a claim is supported once the evidence is pooled. They do not indicate which tool output supports it, nor whether it is indeed the cited source.

The authors take the example of a customer support agent that responds: ‘According to the client file, this package includes a 30-day refund window.’ This duration may be accurate, but appear in a business policy document and not in the cited file. In a clinical agent, information about a patient’s treatment, drawn from a medical history tool, becomes misleading if the response presents it as a finding from the medical literature.

A claim-by-claim check

ProvenanceGuard is applied after the response is generated, treating the agent as a black box, without retraining. The system reads the MCP trace, including tool outputs and their source identifiers, then breaks down the response into precise claims. For each, it identifies the most relevant source, checks if it supports the claim, and compares this source to the one the response names or suggests. It finally produces a verdict per claim and an overall decision, to allow or block the response.

For experiments, the authors used local models: MiniLM to find the relevant source, a natural language inference verifier based on DeBERTa to judge if the source supports the claim, and a local language model to break down the response. The authors specify that these models are not mandatory and that adaptation to cloud-hosted models remains possible, subject to new tests and calibration. A blocked response can be rewritten by a RARR-type repair step, then verified again.

Results

The main test involves 281 traces from a medical agent that consulted patient records and research articles. Human experts examined 361 claims from 40 responses. According to these experts, 139 claims should not pass, and ProvenanceGuard blocked 138. It also set aside for review or repair 67 claims deemed supported, a deliberate choice by the authors in this cautious framework. For claims with an identifiable source, the correct source was selected in about 86% of cases.

On the same set of claims, the blocking F1 score reaches 0.802 for ProvenanceGuard, compared to 0.783 for MiniCheck, 0.758 for RAGAS Faithfulness, 0.662 for AlignScore, and 0.436 for SummaC-ZS. Only ProvenanceGuard associates each claim with its source identifier. In a more difficult test, with multiple similar sources, the blocking F1 score is 0.846, but the exact source is correctly identified in only 50.3% of cases. In another test, 50 misattributions were introduced, and all 50 were detected.

Repair resolved the 173 blocked responses in the full-trace trial, but 144 of them ended with a fallback text rather than a substantive rewrite. On reconstructed multi-source traces, the 59 initially blocked responses were resolved, with only two final fallbacks. In offline use, the overhead is about half a second per response in the tested local configuration.

ProvenanceGuard was also presented as a poster at the Agentic AI Summit 2026 at the University of Berkeley.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x