RESEARCH · RESEARCH · #1195
ProvenanceGuard: source-aware factuality verification for MCP-based LLM agents
The paper ProvenanceGuard proposes a post-generation verifier that preserves source identity from Model Context Protocol (MCP) tool traces to check both whether each claim is supported and whether it is attributed to the correct tool output. The method decomposes answers into claims, finds the most relevant MCP source for each, verifies support (using models in the authors' local setup such as MiniLM and DeBERTa), compares claimed attribution, and can block or prompt source-grounded repairs; experiments used 281 real medical-agent traces.
KEY POINTS
- The paper ProvenanceGuard proposes a post-generation verifier that preserves source identity from Model Context Protocol (MCP) tool traces to check both whether each claim is supported and whether it is attributed to the correct tool output.
- The method decomposes answers into claims, finds the most relevant MCP source for each, verifies support (using models in the authors' local setup such as MiniLM and DeBERTa), compares claimed attribution, and can block or prompt source-grounded repairs; experiments used 281 real medical-agent traces.
- It addresses cross-source conflation—claims supported somewhere in pooled evidence but misattributed to the wrong tool—which is critical for safety and compliance in data-sensitive MCP agent deployments.
WHY IT MATTERS
It addresses cross-source conflation—claims supported somewhere in pooled evidence but misattributed to the wrong tool—which is critical for safety and compliance in data-sensitive MCP agent deployments.