Tool-call ordering failures occupy a category that no amount of prompting fully controls, and practitioners building LangGraph agents in regulated or high-stakes workflows have learned this the expensive way. The developer who open-sourced a custom enforcement layer did so after discovering that a ~95% compliance rate on "call check_policy before issue_refund" is not a workable answer for an auditor asking why a refund went through. LangGraph's Human-in-the-loop mechanism is the official answer for this class of problem, but it does not scale to non-destructive read queries at volume …