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

Biological signals arrive with acquisition metadata that a model interface can silently discard or reinterpret. We study this failure mode in three released EEG foundation encoders. Reordering EEG rows together with their electrode names leaves the physical recording unchanged, so a well-specified interface should either preserve that equivalence or declare the channel order it expects. LaBraM and EEGPT are invariant by construction and at floating-point scale. CBraMod contains a channel-index-tied convolution and is order-sensitive, although pretraining reduces the effect to 0.42× a random-weight control by concentrating 83.1% of channel-axis kernel energy in the centre tap. The more consequential failure is amplitude. All three model papers state an input scale, but none of the released inference paths enforces it. For EEGPT, the maximal electrode-identity perturbation is 252× larger at the declared scale than at raw microvolts; CBraMod’s order effect is 4.85× larger. At declared scale EEGPT’s maximal identity error reaches 0.67× a between-subject representation distance. These hidden contracts also change downstream decisions without producing a simple accuracy warning: one C3/C4 mislabelling changes 11.5% of EEGPT probe predictions, while tenfold under-scaling changes 51.5% and balanced accuracy slightly rises. We propose a six-test conformance suite and three model-card fields for channel identity, required order, and amplitude unit. The failure is not that one architecture chooses a particular convention; it is that released biological-model interfaces can omit conventions that materially determine what the model computes.