ESF: Harmonizing Heterogeneous EEG Acquisition for Pretrained Classifiers

Pretrained EEG classifiers are built on recordings from one setup and then used on recorders that sample, amplify, reference and filter differently. This preprint asks a narrow question: if every recording is first passed through one fixed preprocessing contract, which of those differences stop mattering, for which model, and at what cost? The contract is the Encephlian Standard Format (ESF), six standard preprocessing steps in a fixed order. Five kinds of recorder difference (sampling rate, gain, mains interference, missing or bipolar electrodes, broadband noise) are simulated on raw microvolt recordings of the TUAB evaluation set (n = 276, held out from all training), and three frozen public models (EEGPT, LaBraM, BIOT) are scored with the contract complete and with the one step that addresses each difference switched off. Gain and mains differences are fully undone for every model. A low sampling rate (128 Hz) is undone for EEGPT but not for BIOT or LaBraM; a low-pass sweep shows why: EEGPT ignores everything above 15 Hz, while the other two depend on the 64 to 100 Hz band that low-rate acquisition removes. Missing electrodes and broadband noise are undone for none. Without the contract, two models turn normal recordings into confident abnormal reports and the third turns abnormal recordings into confident normal ones; with it, those failures disappear exactly where the sweep says they should. Permuted-label controls, a label-free monitor that fails, and six recordings from a 125 Hz Natus recorder complete the study. Version 2 corrects version 1: the acquisition operators are now applied upstream of the contract, every probe is trained on the TUAB train split only, and the study is extended from one model to three with four probes each. Code, twelve probes and every per-recording prediction: https://github.com/h5371h/esf-recoverability, tag v2.0.0-preprint-v2, DOI 10.5281/zenodo.22771925. Supplement (proofs, per-cell tables, provenance) is attached.