ProvenanceGuard

Source-Aware Factuality Verification for MCP-Based LLM Agents

Ander Alvarez, Santhiya Rajan, Samuel Mugel, and Román Orús

ProvenanceGuard studies a failure mode in tool-using agents that ordinary factuality checks can miss: a claim may be supported somewhere in the available evidence while being attributed to the wrong source. The proposed verifier keeps source identity explicit throughout claim decomposition, evidence routing, support checking, attribution checking, and answer-level allow/block decisions.

Read on arXiv · Download the repository copy · Cite this work

Key results

• Evaluation on 281 medical-domain MCP-agent traces.
• Block F1 of 0.802 and source accuracy of 0.858 on the held-out split.
• Block F1 of 0.846 on a harder multi-source benchmark.
• Detection of all 50 injected attribution swaps in controlled conflation probes.

These figures are reported by the paper; consult the manuscript for protocols, qualifications, and limitations.

Repository status

This is a public paper-artifact repository for arXiv v2. It includes the manuscript and citation metadata. Reproducible code and data are not currently released, and this repository does not claim otherwise.

Citation

See CITATION.cff or CITATION.bib.

@misc{alvarez2026provenanceguard,
  title         = {ProvenanceGuard: Source-Aware Factuality Verification for MCP-Based LLM Agents},
  author        = {Alvarez, Ander and Rajan, Santhiya and Mugel, Samuel and Orús, Román},
  year          = {2026},
  eprint        = {2606.18037},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  doi           = {10.48550/arXiv.2606.18037},
  url           = {https://arxiv.org/abs/2606.18037}
}

License

The manuscript and preview are available under CC BY-NC-SA 4.0. Citation metadata may be reused freely.