AAGSF Synthetic Authority-State Simulation Dataset and Reproducibility Package
Description
This dataset accompanies the working paper AAGSF: Synchronizing Institutional Authority and Executable Authorization in Autonomous AI Agent Systems. It contains the code, processed simulation outputs, sensitivity-analysis results, and reproducibility documentation supporting the paper’s experimental evaluation of the Autonomous Agent Governance and Security Framework (AAGSF). The experiment is a synthetic scenario-based stress test comparing a conventional Identity and Access Management (IAM) baseline with an AAGSF prototype. The primary experiment consists of 20 independent replicates of 25,000 proposed autonomous-agent actions each, for a total of 500,000 simulated actions. The primary configuration uses a 2% authority-representation error rate and a 2% authority-synchronization error rate. The evaluated scenarios include routine authorized actions, approval-dependent transactions, revoked authority, expired mandates, non-delegable actions, recursive delegation beyond authorized scope, cross-organizational trust failures, jurisdictional restrictions, stale represented authority, and ambiguous authority states. Reported measures include Unauthorized Execution Rate, Authorized Action Block Rate, Revocation Escape Rate, escalation precision and recall, comparative effects, sensitivity to authority-representation error, and a local in-process policy-evaluation microbenchmark. The dataset contains synthetic simulation outputs rather than observational or field data. Scenario frequencies, baseline authorization probabilities, and injected error rates are experimental assumptions and must not be interpreted as estimates of real-world incident prevalence. The synthetic outputs were generated algorithmically by the accompanying Python simulation and can be reproduced using the included source code. No human-subject data, personal data, or confidential organizational data are included.
Files
Steps to reproduce
This dataset contains the reproducibility materials for the AAGSF synthetic authority-state simulation. Install Python 3.x with NumPy and pandas. Download all files in this dataset. Run aagsf_simulation_reproducible.py from a Python environment. The simulation generates synthetic autonomous-agent action scenarios representing authorized actions, approval-dependent actions, revoked authority, expired mandates, non-delegable actions, recursive delegation violations, cross-organizational trust failures, jurisdictional restrictions, stale authority representations, and ambiguous authority conditions. The primary experiment runs 20 independent replicates of 25,000 proposed agent actions each, totaling 500,000 simulated actions. The baseline IAM model evaluates actions primarily using technical permission state. The AAGSF model additionally evaluates institutional authority, represented authority, revocation and synchronization state, delegation constraints, approval requirements, and contextual conditions. Primary outputs include unauthorized execution rate, authorized action block rate, revocation escape rate, escalation precision and recall, scenario-level results, and comparative effects. Sensitivity analysis varies the injected authority-representation error rate from 0% to 10% to evaluate degradation under imperfect authority representation. Generated outputs can be compared with the supplied CSV files to confirm reproducibility. The experiment uses synthetic data only. Scenario frequencies, baseline authorization probabilities, and injected error rates are experimental assumptions and should not be interpreted as estimates of real-world incident prevalence.