Tag
This paper introduces ProvenanceGuard, a source-aware verification method for MCP-based LLM agents to address cross-source conflation, where claims are factually correct but attributed to the wrong source.
The article explains that a cache hit in language models does not prove that computational work was skipped, and introduces LLMTraceFX PR #74, which uses token-granular synthetic controls and multi-check verification to validate cache behavior.
Hedgemony detects when AI-generated code makes false claims by checking propositions against the actual runtime environment, classifying errors like fabrications and contradictions.
The author built OpenPitStop, an open-source tool that independently checks and verifies the work of AI coding agents, demonstrated on a broken application, and invites discussion on how to trust AI changes.