Learning Audit Burden Indices and Measuring Semantic Drift Metrics in Industrial Safety Procedures

A Lock-Out/Tag-Out Pilot Study

Abstract

Lock‑out/tag‑out (LOTO) procedures are critical for isolating hazardous energies before maintenance. AI-assisted authoring tools can help technicians produce an initial draft of these documents, which can then be edited and completed before review, but the resulting procedures still require human validation before deployment. This paper proposes a measurement framework for two complementary quantities: an audit burden index that ranks procedures by expected review effort and semantic drift metrics that quantify how the language of procedures evolves over time. The audit burden index is learned from interpretable content features using a linear latent model trained with operationally derived pairwise comparisons and auxiliary outcome heads. The drift metrics operate on document embeddings and capture shifts in the distribution at different temporal resolutions. We describe the audit workflow in which these measurements arise, present an empirical overview on a cohort of procedures, and summarise pilot results that demonstrate the feasibility of our approach. A detailed mathematical development of the latent model and drift measures is provided, and we discuss training, evaluation and limitations. Throughout we emphasise that the goal is measurement rather than prediction: the index and drift metrics are intended for monitoring and triage, not for automated approval.