EFT–MAST CooperBench Reproducibility Dataset: Frozen Holdout Predictions, Brier Decomposition, and Analysis Code (Phases 8–17)

Overview This dataset provides the reproducibility materials for the exploratory CooperBench evaluation accompanying the manuscript Beyond Agent-Level Alignment: Ethical Field Theory and Defensible Network-Level Governance for Multi-Agent AI Systems. The study examines whether behavioral features inspired by Ethical Field Theory (EFT) provide additional retrospective predictive information about joint task success in multi-agent AI systems. Dataset and methodology Four logistic-regression specifications were trained using 48 labeled CooperBench agent-pair runs from 20 repository/task groups and evaluated using frozen predictions on an independent Qwen3.5-9B Mini-SWE cohort comprising 367 runs from 70 disjoint groups. Only 19 held-out runs (5.18%) succeeded. The archive includes: - Derived numerical features, outcome labels, and frozen predictions. - Cleaned research notebooks covering Phases 8–16. - Analysis code, software requirements, and reproduction instructions. - SHA256 integrity records and prediction-lock metadata. - Group-bootstrap uncertainty estimates and calibration diagnostics. - Phase 17 Brier-score decomposition results. The archive supports independent computational reproduction without redistributing the original full agent trajectories. Main findings The EFT-augmented model achieved a Brier score of 0.51628 compared with 0.52493 for the baseline. The prespecified paired improvement was +0.008652, with a 95% group-bootstrap interval of [−0.000629, +0.018261], which includes zero. A subsequent post-lock decomposition showed that the numerical improvement was more than fully explained by reduced squared mean prediction bias. Both models substantially overpredicted success, and the EFT-augmented model had worse log loss. The findings are exploratory and do not establish the predictive superiority of EFT, validate regulatory awareness or ethical conversion, or demonstrate the effectiveness of proposed governance interventions. Data sources and reproducibility Original benchmark datasets are available from the CooperBench project on Hugging Face: https://huggingface.co/datasets/CooperBench/team-trajectories https://huggingface.co/datasets/CooperBench/qwen9b-coop-mini-swe-agent The held-out dataset revision is pinned in the accompanying metadata. Reproduction instructions are provided in README.md. Cryptographic hashes verify file integrity but do not independently certify the timing of prediction locking. All reported predictions were derived from completed trajectories, making the evaluation retrospective rather than prospective. This archive excludes unrelated synthetic governance experiments and is intended to support transparent methodological evaluation and further independent research.