OrganOS Sigma: Auditing Temporal Fault Detection and Chance-Level Performance in the Tennessee Eastman Process

This preprint presents an evidence-based audit of OrganOS Sigma fault detection on the Tennessee Eastman Process (TEP) benchmark. In a development experiment involving faults 3, 9, and 15, eight residual-based detection methods showed limited discrimination, with the best method exceeding a chance-matched random-alarm baseline by only 1.15 percentage points in F1. In a separate 20-fault evaluation, OrganOS Sigma v8.7 achieved an F1 score of 0.9038 and Youden's J of 0.7841, compared with 0.8764 and 0.7301 for a histogram gradient-boosting baseline. Improvements were concentrated in selected faults, while difficult faults 3, 9, and 15 showed little or no improvement. The study reports class-prevalence effects, chance-matched baselines, fault-specific results, and limitations of run-level detection metrics. Stored experimental evidence was checked using SHA-256 manifests, but independent model reproduction and industrial field validation have not been performed. A public metric-recalculation script is provided. Proprietary OrganOS implementation details, internal architecture parameters, and model weights are not released. This manuscript is a preprint and has not undergone peer review.