Empirical Test (Control Systems): Maneuverability Loss Precedes Failure in a Closed-Loop Process Benchmark

Empirical Result (Read This First) This page shows what coherence loss looks like before failure, using real control systems. What was tested Whether loss of maneuverability in a closed-loop control system appears before observable failure. What was observed Across the tested control-system data, trajectory narrowing and loss of control margin occurred while outputs remained nominal, with observable failure only occurring later. What this means The constraint manifests prior to alarms or threshold violations, providing measurable lead time without model fitting. How to replicate quickly Use any regulated control system with logged state variables. Track option-space contraction (variance / controllability proxy) over time. You should observe trajectory collapse before failure. Data & Method (Details Below) (Leave everything else exactly as it is below this line.) Abstract This project tests whether sustained loss of maneuverability in a regulated control system produces detectable trajectory narrowing prior to observable failure. We use the Tennessee Eastman Process (TEP) benchmark dataset and define maneuverability and control-margin proxies from time-series measurements. We evaluate lead time between proxy degradation and constraint violation under multiple fault types. � Kaggle +2 Dataset Tennessee Eastman Process (TEP) simulation dataset (fault-free + faulty runs; multiple fault classes). � Kaggle +1 Hypothesis In a regulated process, proxy maneuverability (M) and control margin (C) degrade before the process exhibits sustained constraint violation (F). Operational Definitions M (maneuverability proxy): rolling covariance trace (or rolling PCA variance) of selected state variables. C (control margin proxy): rolling indicator of control effort “crowding bounds” (saturation/near-bound persistence), dataset-permitting. T (trajectory narrowing proxy): rolling recovery lag + autocorrelation increase. F (failure): sustained constraint violation (outputs beyond nominal envelope for ≥N samples). Falsification Criteria (decisive) SDTL fails for this domain if, across fault classes: False negatives dominate: failures occur with no prior sustained degradation in M/C/T within the preregistered lead window. Non-specificity: M/C/T degrade equally often in fault-free trajectories without any subsequent constraint violation. Observer dependence: independent scorers applying the same code and thresholds disagree on lead-time detection. Primary Outputs Lead-time distribution per fault class (histogram + table) False positive / false negative rates vs. a naive baseline One canonical figure: M/C/T overlays aligned to first failure time