AXE: Structural Accountability Evaluation Framework for Agentic AI Systems

AXE is a procedural audit framework for evaluating structural governance properties of agentic AI systems. The framework provides: Eight operational primitives (state, update, persistence, authority, observability, irreversibility, optimization pressure, attention) for describing agent structure Five failure families derived from primitive interactions (State, Authority, Observability, Execution, Optimization) Six parametric evaluation templates for probing structural vulnerabilities Adapters for testing Claude, GPT-4o, Gemini, Llama, Qwen, and other models This code accompanies the paper "AXE: Structural Accountability for Agentic AI Systems" submitted to ACM FAccT '26. Installation: pip install -e . Quick start: See examples/ directory Run evaluation: python experiments/run_full_evaluation.py