Reproducibility Dataset for Safety-Oriented Evaluation of Cross-Document LLM Reasoning for Construction Compliance

This dataset contains the reproducibility materials supporting the study “Safety-Oriented Evaluation of Cross-Document LLM Reasoning for Construction Compliance.” The deposit includes synthetic construction-quality evidence packs, requirements, ground-truth annotations, experimental conditions, prompts and methodological materials, model outputs, scoring and analysis resources, source and licence documentation, Python rerun materials, supplementary tables, and supporting documentation. The dataset supports reproducibility of the reported safety-oriented evaluation, including Unsafe Automatic Approval, evidence traceability, escalation and abstention behaviour, and related secondary measures. The evidence packs are synthetic and should not be interpreted as records from live construction projects or contractor submissions.