Benchmarking Reference-Free LLM Agent Robustness Under Schema, Policy, and Toolset Drift: Data and Code Artifact

Episode-level evaluation records, source code (with the vendored tau2-bench subset) and offline verification scripts for the IEEE Access article doi:10.1109/ACCESS.2026.3696096. The records reproduce the article's Tables 1 and 3 exactly when each trial's last record is its outcome (trials cut short by an API credit error were re-run in April 2026 and appended). ERRATA.md lists where the article's Figure 2 and per-drift text differ from the records (E1-E5) and where its description of the method, code and records differs from them (E6-E20); scripts/analysis/ holds the reanalyses, and results/raw_phase_a_2026-03-06/ the calibration run behind the reference-guided rate. Version v0.1.0 held the records as they stood in March 2026, before that re-run.This version (October 2026). All files of version v0.2.0 (DOI 10.5281/zenodo.23586228) are preserved unchanged. The added folder supervisor-2026-10/ holds the analyses, inputs, hashes and limitations from the supervisor review of the PhD thesis; read its README first. It reports no new model evaluation.