EEGT v0.5.0 · Pretrained encoder feasibility

Experiment 009 adds a pinned CodeBrain EEGSSM encoder to EEGT's public research ledger. It preserves the frozen selection and its insufficient-data result.

• 240 candidate 30-second segments from two hours of already exposed ear EEG; 9 segments pass the common quality rules (3.75%, or 4.5 minutes).
• 45 model forward passes: original waves plus four matched waveform controls.
• Zero complete eligible participant pairs. All six primary comparisons report INSUFFICIENT_PARTICIPANTS; no participant-level p values are reported.
• All candidates, exclusions, 27 segment/view comparisons, 36 control comparisons, wave derivatives, embeddings and reproducible code are retained.

The model is a continuous EEG encoder backbone, not a full discrete tokenizer or LLM decoder. Its scalp-to-ear spatial transfer and pretraining overlap remain unvalidated. This release does not establish universal tokens, clinical meaning or physical Neurable compatibility. The source reports 50 Hz mains; the frozen recipe's 60 Hz notch does not specifically remove that component.

Independent review reproduced all 54 segment correlations and 1,431 shifted-time values. A fresh forward agrees within 1e-5 but is not byte-identical. General tests, separate encoder tests, release source bindings and extracted-bundle reproduction are recorded with the release. SHA256SUMS identifies expected download bytes. GitHub Actions is disabled by the organization and is not claimed as passing CI.

Read the Research Note, frozen protocol and reproduction instructions.

The data archive includes nine derived public-source wave blocks, aligned descriptors and all encoder outputs. Full original recordings and pretrained weights remain upstream. SHA256SUMS and the manifest identify the exact release bytes. Earlier releases remain unchanged.