Multi-Agent LLM Classification of STPA Context Tables

This paper presents a multi-agent Large Language Model (LLM) framework that automates the evaluation of STPA context tables, addressing a critical bottleneck in safety analysis workflows. In current practice, safety engineers must manually classify every combination of control actions and process model states as safe or unsafe, a labor-intensive and error-prone task that does not scale to systems with hundreds of state combinations. Existing LLM-based approaches primarily focus on generating Unsafe Control Actions (UCAs) and loss scenarios. Our work focuses instead on this unaddressed middle stage: systematic classification of all context table rows. The proposed framework: (1) automatically derives system understanding from context table structure and configuration data, (2) synthesizes explicit safe/unsafe conditions as formal classification criteria, and (3) performs a two-stage evaluation where initial classifications are re-examined under constraint-based validation by a separate agent. All LLM outputs are structured, schema-validated, and exported to a comprehensive Excel report containing classifications, UCAs, severity levels, and summary metrics. We instantiate and evaluate the framework on a Train Door Control System case study with 48 representative state combinations, demonstrating that multi- stage, self-correcting LLM reasoning can systematically support context table evaluation with traceability suitable for safety-critical applications.