lab Hugging Face Blog · 1d ago · releases

ProvenanceGuard adds source-aware verification to MCP agents

from English

ProvenanceGuard is a post-generation verification layer for Model Context Protocol (MCP) agents that detects cross-source conflation by ensuring claims are attributed to the correct evidence source. The system analyzes captured MCP traces without retraining the agent, breaking answers into claims and verifying that the supporting source matches the one named or implied in the output.

ProvenanceGuard caught 138 of 139 claims that human experts deemed unsupported while correctly identifying the right source about 86% of the time.
In a test with similar sources, it scored 0.846 F1 for blocking decisions and identified exact sources in 50.3% of cases.
It successfully detected all 50 instances of wrong attribution where the named source was swapped while evidence remained intact.
The system integrates a RARR-style repair loop that resolved all blocked answers, though many resulted in safe fallback text rather than substantive rewrites.

This approach addresses the limitation of faithfulness scores by making provenance visible claim-by-claim, which is critical for data-sensitive settings where wrong attribution can be as damaging as incorrect facts.

Importance 1/3 New harness with differentiators Trust 1/3 Hugging Face Blog Retrieval & RAG Safety & alignment

Related articles

media Hugging Face Forums · 2d ago · 2 views

Grok connects Covenant project to Robert Hanson's work without being asked

The author of the peer-to-peer ledger "Covenant" reports that Grok on X independently compared their project to the work of Robert "Grizzly" Hanson, despite no prompt or prior mention. This marks the third instance where an AI system has linked Covenant to external work without being requested.

arxiv arXiv cs.CL · 44d ago · 35 views

Ingestion-time defenses fail against coordinated vector retrieval poisoning

A study demonstrates that admission-time filters cannot prevent coordinated poisoning attacks on vector retrieval systems, as geometrically identical patterns exist for both legitimate niche uploads and malicious injections. The researchers show that a small number of unremarkable documents can surround a target query to seize top-k results in BGE-large / BEIR benchmarks.

media Hugging Face Forums · 65d ago · 11 views

DESi Review: an installable Claude skill for structured epistemic paper review

The author has released DESi Review, a practical application of the model-independent DESi epistemic governance framework packaged as an installable Claude skill. This tool runs a structured paper-review pipeline over English text by separating deterministic claim extraction from model-based adjudication.

arxiv arXiv cs.CL · 77d ago · 15 views

PCC-SQL enforces column-level access control in Text-to-SQL via policy-conditioned constrained decoding

Researchers introduce PCC-SQL, a system that integrates column-use policies into text-to-SQL generation to ensure deterministic compliance with data access rules. The approach formalizes policies based on semantic use—output, filter condition, and aggregation argument—and aligns them with grammar productions tracked by the decoder.

arxiv arXiv cs.CL · 79d ago · 12 views

Study finds RAG transfers ideological discourses to LLM outputs, affected by temperature

A study examines how Retrieval-Augmented Generation (RAG) frameworks transmit ideological bias from retrieved materials into large language model (LLM) responses. The research identifies that sampling temperature significantly influences the strength of this ideological transfer.