Equivalent EEG, Different Representation: Undeclared Acquisition Contracts in EEG Foundation Models

An EEG recording is stored as one waveform per electrode. If we reorder those rows and carry the electrode names with them, the physical recording has not changed. We use that simple equivalence to audit three released EEG foundation models: LaBraM, EEGPT, and CBraMod. Two models are invariant to the joint reordering by construction, and their measured differences are only floating-point noise. CBraMod contains a convolution tied to channel index and is order-sensitive, but pretraining reduces the effect: 83.1% of its channel-axis kernel energy moves to the centre tap, and the released checkpoint shows 0.42× the order dependence of a random-weight control. A second issue matters more. Each model paper states an input amplitude convention, yet the released inference paths do not enforce it. For EEGPT, the maximal electrode-identity perturbation is 252× larger at the declared scale than at raw microvolts; for CBraMod, the order effect is 4.85× larger. At declared scale EEGPT’s maximal identity error reaches 0.67× a between-subject representation distance. We then ask whether these representation changes affect decisions. A single C3/C4 mislabelling changes 11.5% of EEGPT linear-probe predictions, and tenfold under-scaling changes 51.5%, even though aggregate accuracy slightly rises. The paper turns these observations into a small conformance suite and three model-card fields. The main lesson is methodological: before comparing frozen representations, make the model’s channel and amplitude contracts explicit.