# ProvenanceGuard adds source‑aware verification to multi‑tool LLM agents

> Hugging Face’s new ProvenanceGuard layer checks that each claim an agent makes is backed by the exact source it cites, catching mis‑attributions in medical and other domains.

Oossa · 2026-09-29 · https://oossa.com/en/provenanceguard-adds-source-aware-verification-to-multi-tool-llm-agents

Hugging Face released ProvenanceGuard, a post‑generation checker for Model Context Protocol (MCP) agents. It keeps the IDs of tool outputs alive, matches each claim to its true source, and blocks answers that attribute facts to the wrong source. In a medical test on 281 agent traces, human reviewers flagged 139 claims as wrong; ProvenanceGuard caught 138 of them and correctly identified the source 86% of the time.

## The facts

- 138 of 139 mis‑attributed claims were blocked in the medical evaluation
- Source‑identification accuracy was 86% on the same test

## Why it matters

Ensuring the right source backs each claim helps prevent misleading information in sensitive fields like healthcare and finance.

## Sources & references

1. [Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents](https://huggingface.co/blog/MultiverseComputingCAI/getting-the-source-right-not-just-the-fact-source) – Hugging Face, 2026-09-29

Last updated: 2026-09-29