Deterministic ToolMisuseBench reproduction artifacts for Detecting, Verifying, and Recovering from Tool Failures in LLM Agents

This record supports Figure 2 and Section 7.3 of a critical review of tool failures in LLM agents. It contains the pinned protocol, offline runner, per-task scores and source values for two deterministic ToolMisuseBench reference agents evaluated on the same 1,200 public-test tasks. The score files were generated by rerunning the two agents. Per-task metrics matched the results reported by ToolMisuseBench. Task identifiers and fault labels come from ToolMisuseBench by Akshey Sigdel and Rista Baral. The pinned upstream code and data are linked in the README and retain their MIT terms. CC BY 4.0 applies to the new reproduction materials, subject to the upstream MIT terms for material drawn from ToolMisuseBench.