Evidence-grounded road safety countermeasure recommendation from crash-context queries: A modular LLM framework

Traffic safety countermeasure recommendation requires simultaneous consideration of crash context, roadway facility, traffic control, operating conditions, and evidence applicability. Traditional rule-based and keyword-matching methods have difficulty processing natural-language crash-context queries. This study proposes an evidence-grounded and mechanism-aware framework that formulates traffic safety countermeasure recommendation as the construction, selection, and ranking of CMF Evidence Cards. A structured CMF Evidence Store containing 8844 Evidence Cards was developed from the CMF Clearinghouse. A corpus of 2935 crash-context queries was further constructed, including 1935 complete queries and 1000 controlled incomplete variants. For each query, the framework constructs a candidate evidence pool through Context-aware Retrieval, Mechanism-aware Recall Augmentation, and Engineering Applicability Filtering. A large language model fine-tuned through supervised learning then selects and ranks up to three Evidence IDs and returns the corresponding countermeasure descriptions, CMF or CRF values, evidence quality, applicability conditions, and literature sources. The fine-tuned Qwen3-32B achieved an Exact Evidence Top-1 of 53.14%, an Exact Evidence Hit@3 of 76.15%, a Mechanism Hit@3 of 83.26%, an Exact Evidence Jaccard of 44.18%, and an Empty-set Decision Accuracy of 95.08%. Module-level ablation results confirmed the distinct contributions of the three candidate evidence construction components. Candidate-pool sensitivity analysis showed that evidence coverage reached saturation at approximately 30 Evidence Cards. The proposed framework also substantially outperformed a conventional BM25-RAG baseline and maintained stable performance across semantically equivalent query paraphrases. LLM-assisted and independent expert evaluations further supported the contextual relevance, mechanism consistency, engineering applicability, and overall usefulness of the recommendations. Case studies demonstrated the framework's ability to address mechanism-complex and facility-specific traffic safety scenarios, while also identifying evidence retrieval as a limitation in specialized contexts such as work zones.