This repository accompanies a preregistered secondary-data study on latent autonomic regulation states in physiological time series. The project investigates whether multichannel physiological recordings are better described by switching state-space models with recurrent latent states than by a single non-switching baseline, and whether such states show interpretable fingerprints, partial cross-dataset invariance, and frozen-definition generalization across independent datasets. Three datasets (WESAD, CASE, and VitaStress) are analyzed using a common pipeline based on protocol-grounded baseline–task–recovery segmentation, fixed 1-second windows, within-dataset standardization, and candidate solutions K = 2–10 compared against a K = 1 baseline. The repository contains analysis scripts for all hypothesis levels (H1 existence, H2 fingerprints, H3 invariance, H4 generalization), derived windowed data, model-fit summaries, fingerprint exports, and implementation notes documenting deviations from the original preregistered dataset plan.