September 29, 2026
MCP agents will be able to verify their answer sources: ProvenanceGuard reaches 0.858
Across 40 medical MCP traces, ProvenanceGuard achieved source-selection accuracy of 0.858 for 260 claims (preprint, Aug 27, 2026). The method breaks a response into individual claims, matches them against tool results, and blocks those that do not match the selected source.
Source-blind methods could verify a claim, but did not tie it to a specific tool result. ProvenanceGuard stores the full tool output together with `tool_name` and `source_id`, so an agent can check not only the answer but also its provenance.
Checking the trace. In 50 controlled attribution swaps, the method blocked every response, and after repair, no incorrect attribution remained. In a difficult multi-source test, accuracy for selecting the “source plus relation” pair fell to 0.229: similar tool results are still hard to distinguish.
In a more difficult multi-source test, block F1 was 0.846.
