Retrieval-Augmented Contract Intelligence for Performance-Driven Construction Governance and Smart Contract Synthesis

Abstract

Construction projects rely on extensive contract documents to govern payment, scheduling, change management, risk allocation, and dispute resolution. The scale and heterogeneity of these documents make contract administration largely manual, increasing the likelihood of misinterpretation, delayed actions, and governance inefficiencies. While recent advances in large language models have enabled automated clause extraction and classification, existing approaches remain disconnected from project performance metrics, standard-form contract requirements, and executable contract logic. This paper presents a retrieval-augmented contract intelligence system for end-to-end construction contract analysis and structured smart contract synthesis. The framework integrates clause extraction, semantic classification, KPI-aware importance ranking, standards alignment, discrepancy diagnostics, and constrained logic synthesis. Clause importance is quantified by mapping contract language to four key project performance indicators: cost overrun impact, schedule delay impact, cash-flow adequacy, and dispute frequency. The system was evaluated using five executed contracts from a large public owner representing general contractor, construction manager, architect-engineer, commissioning, and design-build delivery methods, comprising 588 clauses. Classification performance on a manually labeled validation subset achieved a macro-averaged F1 score of 0.55, with inter-annotator agreement of 69.17% (Cohen’s κ = 0.52). Clause prioritization rankings remained highly stable across alternative KPI weighting scenarios (Spearman ρ > 0.98). A contract-context audit further refined standards-based missing-provision findings by distinguishing confirmed omissions from relocated or uncertain obligations. For automation outputs, constrained template-based synthesis improved smart contract quality scores from 0.8/4.0 to 3.6/4.0 relative to unconstrained generation, while execution-level validation achieved a 100% pass rate across representative trigger scenarios. The findings demonstrate that retrieval-augmented language models, when combined with structured performance reasoning and standards-aware controls, provide a scalable and explainable foundation for smart contract-enabled contract governance in construction projects.