Reproducibility Package for "A Governance Architecture for AI-Assisted Validation of Construction Cost Knowledge: Combining Domain Rules with LLM Semantic Analysis"

Reproducibility package for the paper "A Governance Architecture for AI-Assisted Validation of Construction Cost Knowledge: Combining Domain Rules with LLM Semantic Analysis" submitted to the Journal of Construction Engineering and Management (JCEM). Contents: Sanitized validation datasets (expert-labeled n=100 sample, portfolio-scale n=1,025 four-way ablation results, baseline ladder outputs, and inter-rater reliability data) Single-command verification script (verify_all_paper_numbers.py) that reproduces all table-level metrics derivable from the released data, including expert agreement metrics (F1, precision, recall, Cohen's kappa), confusion matrices, cross-region analysis, baseline comparisons (LR-S2 via GroupKFold), and risk distribution across validation modes Frozen official McNemar and bootstrap delta-F1 artifacts from the original production evaluation LLM prompt template and structured output schema used for semantic analysis Reproducibility supplement (PDF) Sanitization: Raw BOQ text, unit prices, and tender names are blanked for commercial confidentiality. Identifiers (FormulaId, TenderId) are pseudonymized consistently across all files. Component percentage features and derived risk classifications are preserved in full. Usage: python verify_all_paper_numbers.py --data-root research/data --output-dir outputs Requirements: Python 3.9+, pandas, numpy, scikit-learn, scipy (see requirements.txt)