ProvenanceGuard: source-aware verification for MCP agents
ProvenanceGuard is a post-generation verifier that preserves MCP tool/source IDs, checks per-claim support and attribution, and issues per-claim verdicts plus an answer-level allow/block decision.
What changed
Adds a source-aware verification layer (ProvenanceGuard) that preserves source identity through claim decomposition, routing, support scoring, attribution checking, and repair, producing per-claim source verdicts and an answer-level allow/block decision.
Why it matters
Developers and reviewers can detect claims attributed to the wrong tool output, reducing cross-source conflation risk in data-sensitive settings like medical records and account policies.
Who should care
DevelopersResearchersEnterprises
Worth trying
Yes — Paper and results show an offline local setup that teams can adapt.