AXLminds has added an experimental EEGNet workflow for studying learned EEG decoding alongside the existing signal-processing approach. It includes dataset checks, separate development and test handling, model loading, replay, and observation-only live logging. The reviewed integration remains on an unmerged research branch.
Synthetic fixtures exercised training, checkpoint loading, and the recording pipeline. A model trained on synthetic data was also replayed over saved recordings and used in a separate 30-second worn observation run on 24 September. This checks that recorded signal can reach the model and the logger. Its synthetic classes have no participant-task meaning; the outputs are not commands or an accuracy result.
The reviewed inventory established no eligible task-training dataset. No task-trained EEGNet checkpoint, real held-out EEGNet score, or matched decoder comparison is available. Task definitions, success targets, timing rules, and reserved-test registration still require completion before confirmatory evaluation.
The latest documented full software retest on 25 September had three synthetic worker deadline failures, although that worker module and the focused guided-session checks passed separately. An earlier revision passed its full suite, and a separate dependency-enabled selection passed 380 tests. Those earlier results do not close the revised full-suite gate, remote review, or physical validation.
Progress now requires verified labels on independent worn sessions, frozen evaluation rules, real training and held-out comparison, and a separate observation run with the resulting model. Software wiring checks cannot substitute for any of those steps.