Start with the signal
An EEG system begins with measurements, not meanings. Before designing an interaction, we need a clear description of the signal: channel layout, sampling rate, time window, and preprocessing. A model is only one component of that contract.
Reconstruction is not decoding
ZUNA and ZUNA1.1 are upstream EEG reconstruction models by Zyphra. Oraculum’s Core ML conversions package that behavior for Apple platforms. Denoising a recording or estimating a missing channel does not establish that a system understands a user’s intended command. A reconstructed signal is still a model estimate.
Different representations, different roles
The CBraMod embedding package produces features for downstream work. The separate motor-imagery classifier produces left/right outputs under a specific cue-paced protocol. These artifacts have different outputs and should not be treated as interchangeable. Good systems preserve those distinctions.
From a benchmark to an interface
A consumer-compatible channel layout is not the same as validation on a live consumer headset. Timing, hardware, artifacts, and the absence of explicit cues can all change the problem. The next experiment should test the actual interaction being proposed, rather than extrapolate a broader capability from a narrow result.
Source material
This editorial draft draws on the public release documentation below. It does not report a new experiment.
ZUNA1.1 Core ML model cardCBraMod embedding model cardMotor-imagery model cardPrepared for this website on 25 Sep 2026.