Tag
This paper proposes a provenance-based framework and multi-stage pipeline, \tool, to detect misalignment in LLM agents' tool invocations before execution, reducing error rates significantly compared to LLM-as-a-judge baselines.