A governance audit framework for large language model advice in civil infrastructure decision-making illustrated with local risk models in Sugar Land, Texas

Abstract Large language models are beginning to appear in civil engineering practice, yet their advice often reflects generalized narratives rather than local evidence. This perspective introduces a governance audit framework that compares LLM outputs with risk models built from local infrastructure data. The approach is designed to help engineers and city managers evaluate whether generative systems emphasize the same drivers that empirical models identify. The audit framework involves four steps: building a local risk model, preparing a concise factor card, running large language models with and without context, and comparing the emphasis of their responses with planning outcomes. The method is illustrated with water main data from Sugar Land, Texas, where a local XGBoost model trained on 35,508 pipe segments (ROC-AUC = 0.909) identified pipe length and age as the dominant failure predictors. Two commercial LLMs overemphasized material by a factor of 2.2 and underemphasized length by a factor of 3.2 relative to their empirical importance. Web search did not meaningfully reduce this divergence. This misalignment underscores the risk of misplaced emphasis and highlights the need for structured audits before adopting generative systems in infrastructure planning. The Perspective also discusses how physics-informed models and digital twins can be integrated with audits to strengthen governance, and outlines a research agenda that includes temporal validation, cross-domain generalization, retrieval-augmented generation, and multi-city studies. The goal is to provide civil engineers with a reproducible framework that ensures generative systems are aligned with local evidence and professional oversight.