Urban rail construction is one of the most safety-critical engineering domains, where even minor errors can lead to catastrophic consequences. Large Language Models (LLMs) offer promising capabilities for automating safety risk extraction from accident reports; however, their deployment is critically hindered by hallucinations – outputs that are fluent yet factually incorrect or logically inconsistent. Existing prompting methods such as Chain-of-Thought rely on implicit reasoning without explicit problem representation, leading to brittleness, logical drift, and undetected constraint violations. This paper addresses key research gaps in current approaches, including the lack of constraint grounding in authoritative knowledge bases, the absence of explicit validation of reasoning consistency, and the limited ability to produce structured, machine-readable outputs. To address these challenges, we propose a conceptual framework, MFR-RAG-V (Model-First Reasoning with Retrieval-Augmented Validation), which separates problem modelling from text generation and enforces structured reasoning through three mechanisms: RAG-based constraint grounding, JSON schema enforcement, and model-based validation using an LLM-as-a-judge. Through a design-oriented evaluation, we illustrate how MFR-RAG-V is designed to provide systematic mechanisms, both preventive and detection-oriented mechanisms, for addressing factual and logical hallucinations. A conceptual comparison with existing approaches suggests that the framework offers theoretical mechanisms that support traceability, consistency, and validation. The proposed framework contributes a structured reasoning paradigm for LLM-based safety applications and provides a foundation for future empirical validation in construction safety management.
Phung et al. (Mon,) studied this question.