Proposing an LLM-Based Framework for Automated Structuring and Verification of Korean Railway Software Safety Requirements

Abstract

Railway signaling software must satisfy strict requirements—clarity, completeness, verifiability, and traceability—from the specification phase. KRS-SG-0069 requirements are written in Korean natural language, making manual structuring costly and expert-dependent. I propose a semi-automated framework combining LLM-based structuring with deterministic verification. Analysis of 94 requirements identified seven defect types mapped to EN 50128 clause 7.2.4 violations. The framework comprises four modules: M1 classifies requirements; M2 converts them into a six-element boilerplate via Chain-of-Thought prompting; M3 validates drafts through a four-layer rule engine; M4 generates state transition models, test cases, and a traceability matrix. The LLM is confined to draft generation, while safety judgments are delegated to deterministic engines and formal validators. This paper presents an architectural foundation; quantitative evaluation and broader applicability remain as future work.