Decoding What You Imagine from What Others Saw

Decoding visual imagery conventionally requires imagery data from the end user, which is slow, effortful, and the bottleneck for deployable brain-computer interfaces. We ask whether that requirement can be dropped entirely by training on perception in other people. On a public EEG dataset of 22 participants, a standardised posterior alpha-power decoder trained on the pooled perception epochs of 21 participants separates the imagery epochs of the held-out 22nd at 5.3 points above a matched permutation null on objects (p = 0.0002, 18 of 22 participants), with figures moving in the same direction as directional support. A pre-declared negative control sits at chance while its perception control is strong, localising the failure to the perception-to-imagery crossing rather than the decoder. Because cross-participant, cross-modality claims admit several routes to above-chance accuracy that involve no shared neural representation, we specify and report the controls that exclude them: presentation-order leakage, cue-locked temporal leakage, ocular contamination, and a 205-configuration model sweep. Transfer is carried by per-participant alignment, not model capacity. We offer these controls as the minimum standard for transfer claims of this kind.