Brain-computer interface feasibility for YazSes — 2026-11-03¶

Status: research and design snapshot, not a released capability, clinical claim, approved human study or permission to alter production.
Tracking: #696 (EEG/EOG feasibility; closed 2026-11-03) under #692 (multimodal HCI); complementary research under #790 (accessible hands-free workspace), especially #814 (study/review gates), #815 (multimodal arbitration), #816 (stop/recovery), and #817 (task evaluation).
Date and evidence: 2026-11-03. Third-party findings below are cited to upstream scientific articles and library documentation; no new device, human, or EEG-model benchmark was run for this report.
Scope: ADR-030 (Accepted); ADR-v2-152 (Proposed, not yet binding); biosignal validation protocol and HCI lab roadmap.

Executive recommendation¶

Investigate non-invasive scalp EEG as a very low-bandwidth intentional switch: one deliberate event versus rest, optionally combined with YazSes voice and gaze. Do not describe muscle or ocular artifacts on an EEG headset as a decoded thought. Do not pursue free-form consumer EEG thought-to-text as an integration promise in 2026. Carefully distinguish synthetic feasibility, published offline dataset accuracy, real-time detection, and everyday assistive usability.

Decision now: (1) keep the existing, off-by-default EMG/EOG/blink/jaw-clench pipeline as the practical control baseline; (2) perform a research-only neural motor-imagery (MI) rest-versus-activate offline replay evaluation when lawful datasets and channel montage permit; (3) require robust rejection of unrelated states, false-trigger exposure estimates and repeated-session evidence before considering a live neural input adapter; (4) leave P300/SSVEP stimuli and invasive speech BCIs out of product scope. No new src/ entrypoint, model download, network listener, runtime feature, public dataset of user signals or hardware capability claim results from this report.

1. What exactly is being decoded?¶

The currently shipped YazSes biosignal switch uses artifacts, not motor-imagery brain decoding: user guide and biosignal spec. The explicit exclusion of neural motor imagery and cortical decoders is recorded in ADR-v2-152, which is Proposed, so it is not yet a binding decision. BrainFlow hardware acquisition support alone does not imply that the app decodes thoughts.

2. Important findings available by 2026-11-03¶

Non-invasive language decoding is advancing but remains limited. The 2026 Nature Neuroscience Brain2Qwerty study tested 35 volunteers typing briefly memorized sentences. Mean character-error rate was 65% on EEG and 29% on MEG (lower is better), with 18% in the best participants on the more informative modality. These are sentence-production during overt keyboard typing experiments with dedicated equipment, not a consumer headset silently reconstructing arbitrary internal speech. Lévy et al., 2026. Do not quote the best-case score as a generic YazSes or consumer-EEG performance estimate.

Motor imagery EEG still struggles with user/session variability. A 2026 review identifies low signal-to-noise, cross-subject/cross-session variability and nonstationarity as persistent hurdles for practical motor-imagery BCIs. Wang et al., Frontiers in Neuroscience, 2026-09-08. An October 2026 Journal of Neural Engineering study specifically reports difficulty rejecting unfamiliar movement classes using uncertainty estimates, showing that generic confidence/OOD gating cannot be assumed to work for all users. Mulder et al., 2026-10-09.

Invasive speech BCIs are a different class of technology. A 2025 Nature Neuroscience clinical-trial study streamed brain-to-voice synthesis from high-density cortical recordings in a participant with severe paralysis. Littlejohn et al., 2025. A separate implant study reported a voice-synthesis interface using 256 intracortical microelectrodes. Wairagkar et al., Nature, 2025. A June 2026 Nature Medicine single-participant study demonstrated nearly two years of independent home use of an intracortical BCI for speech/cursor control. Card et al., 2026-06-15. These establish real clinical research progress, not a general-purpose non-invasive brain-to-text decoder.

Open, reproducible tooling exists. BrainFlow's supported-board documentation documents a synthetic board with no actual hardware, and a common acquisition API for supported real boards. MOABB's documentation provides EEG benchmark pipelines and stresses reproducibility; its 2026 v1.6.0 update added imagined-speech dataset categories. MOABB v1.6.0 release notes (Zenodo record) and v1.7.1 (Zenodo record) document versions, but dataset rights are separately determined. MNE-Python EEGBCI and the PhysioNet EEG Motor Movement/Imagery dataset are potential subject-level research resources, subject to review of exact dataset terms.

Visual stimulation is not a neutral accessibility option. A 2025/2026 Epilepsy Foundation working-group consensus identifies hazards from certain bright flickering patterns and recommends caution for visual media. Fisher et al., Epilepsia, 2026. Therefore no P300/SSVEP flashing prototype or participant exposure is proposed by this document; any future study needs a specialized safety and accessibility review.

3. Reuse the existing YazSes architecture¶

There is no reason to build a second biosignal acquisition or activation system. The already documented implementation encompasses:

• yazses.biosignal: quality gating for missing/flat/saturated/noisy signals; sample dropout and clock integrity; conservative blink, clench and squeeze events; cancellation / false-positive governor; optional BrainFlow, LSL, local replay and synthetic sources. See v2.48.0 release evidence. Its device paths were tested with fakes/fixtures, not real hardware.
• platform/emg/ and activation-source seam: existing physical-switch event routing, not a cortical decoder.
• design/hci-huggingface/ research tasks #692/#696/#699/#700: shared synthetic samples, offline replay, uncertainty/abstention/confirmation tests.
• design/eye-control/ research and review: target semantics, safety, consent and device testing.
• biosignal field protocol: baseline measured phantom activations/hour and hit rates, not an ambiguous classifier 'accuracy'.

A hypothetical separate research adapter could output candidate events, not commands, with this conceptual-only envelope:

NeuralActivationCandidate (RESEARCH ONLY)
  source_kind: "eeg_motor_imagery"      # NOT "blink" or "jaw"
  candidate_kind: "activate" | "rest" | "abstain"
  timestamp_monotonic_ns: integer
  confidence: float                   # calibrated estimate, not assumed accuracy
  quality: "valid" | "invalid"
  model_revision: immutable identifier
  calibration_session: local opaque identifier
  expires_after_ms: bounded integer

Hard rules: a candidate does not type, click, open the mic, call an agent, or send text. Missing channel quality, old timestamps, clipped frames, unrecognized state or low confidence produces abstain. No activation across restart without explicit opt-in; user controls pause, kill, re-enable. No recordings or derived neural features in text logs. This is a design proposal only; it is not a new production wire format and does not amend ADR-030/ADR-v2-152.

4. A feasible staged, falsifiable research protocol¶

Stage R0 — inventory and synthetic signals (no people, hardware, or model download)¶

1. Audit existing biosignal, activation, BrainFlow, dataset schemas and #696, #699, #700 before adding files.
2. Specify distinct metadata tags for neural EEG, EOG artifact and EMG artifact; never recast a blink as brain decoding.
3. Replay deterministic synthetic scenarios with timestamp jitter, channel dropout, out-of-order samples, flatline, saturation, mains interference, natural blink/jaw movement, stale model scores and stream interruption.
4. Score quality-gate abstention, stop/recovery invariants and false-positive accounting without executing desktop actions.
5. Important limitation: BrainFlow synthetic board samples are valid for testing acquisition plumbing and failure handling; they cannot prove motor imagery decoding or user intent.

R0 exit: 100% reproducible no-hardware, no-egress event/abstention tests; no backend added.

Stage R1 — offline public-dataset neural study (after provenance review)¶

Use an appropriately licensed EEG motor imagery dataset with documented motor-cortex channels, task labels, rest trials, sample rate, timestamps, participant boundaries and repeated sessions. A generic frontal consumer EEG band must not be assumed adequate for motor imagery. Freeze dataset id/version/license, inclusion/exclusion, participant grouping and normalization before model selection; never mix windows from the same contiguous recording across train/test splits.

Suggested baselines (compare, do not assert superiority):

• Classical bandpower and common spatial patterns (CSP) + shrinkage LDA or regularized logistic regression.
• Riemannian covariance methods for MI feature comparison.
• Lightweight EEGNet-style classifier only as an explicitly justified candidate, with parameter and compute parity noted.
• Negative controls: shuffled event labels, signal-only artifacts, no-intent rest, unfamiliar movements, speaking, blink/jaw activity and channel dropout.

Evaluate within-participant held-out sessions, and separately leave-one-participant-out transfer. Treat any absence of repeated sessions as an inconclusive generalization study, not evidence of a ready-to-use interface. Report data leakage checks, imbalance, held-out failure cases and class-conditional abstentions.

Stage R2 — streaming pseudo-online replay (no live UI actions)¶

Turn offline EEG windows into a causally ordered stream. Fit any normalization and calibration only on past training data. Require a predeclared decision threshold and refractory period; never compute future samples when scoring latency. Evaluate:

• True intentional triggers per labelled cue; missed activations / recall and precision.
• False activations per usable idle hour, including negative controls for speech, natural blinking and head motion. Report a one-sided 95% upper bound (Poisson model with assumptions disclosed); zero events in a short test is not proof of safety.
• Time from cue/onset to confirmed stable candidate (median and p95), including sensor buffering and inference.
• Calibration time and degradation from one session to the next; per-person failures, not just pooled means.
• Unknown/OOD state detection, threshold sensitivity, signal quality failure rate, abstention rate and false "activation" during disconnection.
• CPU and peak RSS on a normal Linux laptop and impact on simultaneous YazSes dictation.

A nominally high within-epoch classification accuracy is not an acceptable replacement for continuous idle exposure. Report an always-rest and random/shuffled baseline. Pre-register the acceptable error/false-trigger budget, then estimate the amount of idle exposure needed to substantiate it; do not choose thresholds after looking at final test data.

Stage R3 — future hardware and human feasibility gate (NOT AUTHORIZED)¶

Requires (a) demand evidence, (b) explicit ADR-030/ADR-v2-152 scope decision, © owner-approved hardware and license/privacy review, (d) proper institutional ethics/review determination where applicable, accessible consent and withdrawal, (e) a participant safety and stopping protocol, (f) an independent desktop test with actions still disabled at first. The human-study process belongs to #425, #427, #814 and the proposed research-consent guardrails in ADR-v2-150, not to a new informal participant pool.

Do not collect diagnoses, psychological or cognitive traits; do not upload raw EEG, EMG, recordings, biometric features or calibration profiles. Explicit public opt-in is not automatically informed research consent. Any research on neural signals should have an appropriate security and retention review.

5. Three design alternatives and decision¶

Stopping conditions (research): absent adequate EEG channels/labels or dataset rights; unsupported/offline-unavailable model; detectable label leakage; refusal/abstain mechanism fails on missing/unknown signal; poor across-session generalization; false-activation upper confidence bound remains unacceptably high; comfort/calibration effort outweighs ordinary assistive input; or any data transfer incompatible with ADR-011/019. Report negative findings and close the candidate, not just tune until the numbers look good.

6. Provenance and reproducibility fields¶

Each research run should record: source dataset ID and URL; rights/provenance and explicitly permitted use; immutable dataset version and local SHA-256; participants and sessions represented (pseudonymous grouping only); montage, sampling rate, signal modality, event labels, preprocessing steps, epoching/causality, sample exclusions; algorithm/library versions and exact full git revisions, seed, hyperparameter protocol, window split rules, confidence calibration, threshold and refractory period; offline-only dependency/egress check; device hardware/OS; metrics and uncertainty for all negatives. Do not commit participant-level signal files or person-identifiable metadata.

For reproducible software examples use BrainFlow synthetic board, a supported MNE pipeline and vetted MOABB datasets. Downloading a package or dataset is a separate step requiring a lawful provenance and license decision. If a scientific paper is later written, use public aggregate tables and code where licenses permit; determine authorship, data permissions, ethics and declarations independently.

7. Suggested project-specific sequence¶

1. #699 (closed) / #696 (closed): both issues are closed, so this review is a follow-up that needs a maintainer decision to reopen one of them or to open a new issue: review the shared synthetic fixture contract and publish documentation-only neural-versus-artifact tag definitions. No forked schema.
2. #696 (closed; same follow-up caveat): specify causal rest-versus-MI evaluation design and public-dataset eligibility; use MOABB's existing benchmark conventions where appropriate.
3. #817 / #816 / #815: agree task-success, false-trigger, abstention and kill-switch metrics applicable to an eventual hands-free interface.
4. #814 / #425 / #427: determine participant research requirements before any live testing.
5. #819 / #790: after research evidence exists, prepare a narrow proposed ADR only if continued work is justified. Record the reason to reject as readily as a reason to proceed.

This document makes no claim of mind-reading, medical validity, generalizable BCI performance or user-tested functionality. It does not create a new delivery task or alter approved product direction.

Selected primary sources (as at 2026-11-03)¶

1. Lévy J. et al. (2026-06-29). Noninvasive decoding of typed sentences from human brain activity. Nature Neuroscience. https://www.nature.com/articles/s41593-026-02303-2
2. Wang L. et al. (2026-09-08). Cross-variability decoding for motor imagery EEG signals: a comprehensive review. Frontiers in Neuroscience. https://doi.org/10.3389/fnins.2026.1899645
3. Mulder M.Q. et al. (2026-10-09). The challenge of out-of-distribution detection in motor imagery BCIs. Journal of Neural Engineering. https://pubmed.ncbi.nlm.nih.gov/42790484/
4. Card N.S. et al. (2026-06-15). Long-term independent use of an intracortical brain-computer interface for speech and cursor control. Nature Medicine. https://doi.org/10.1038/s41591-026-04414-6
5. Littlejohn K.T. et al. (2025-03-31). A streaming brain-to-voice neuroprosthesis to restore naturalistic communication. Nature Neuroscience. https://doi.org/10.1038/s41593-025-01905-6
6. Wairagkar M. et al. (2025-06-12). An instantaneous voice-synthesis neuroprosthesis. Nature. https://doi.org/10.1038/s41586-025-09127-3
7. Fisher R.S. et al. (2026-02). Visually-provoked seizures: Consensus of the Epilepsy Foundation Working Group. Epilepsia. https://pubmed.ncbi.nlm.nih.gov/41283196/
8. BrainFlow. Supported boards and synthetic board. https://brainflow.readthedocs.io/en/stable/SupportedBoards.html
9. MOABB. Quickstart and reproducible benchmark mission. https://moabb.neurotechx.com/docs/README.html ; release v1.7.1 https://github.com/NeuroTechX/moabb/releases/tag/v1.7.1 (Zenodo record: https://zenodo.org/records/22215752)
10. PhysioNet. EEG Motor Movement/Imagery Dataset. https://physionet.org/content/eegmmidb/1.0.0/

Update: executable synthetic motor-imagery negative control (2026-11-03)¶

The repo already had a tested P0/P1 synthetic EOG/EMG artifact replay in the HCI protocol. To avoid duplicating it, this research branch adds a separate, standalone synthetic motor-imagery stress experiment, script and 13 regression tests.

Key negative finding (authored signals only): the toy 10 Hz power detector generated two proposals for two deliberately easy imagery blocks, but also one false candidate on a rest interval with lower synthetic baseline power. Blink, muscle, missing-data and flatline scenarios produced no proposals in that seed. This is not a human EEG result; all data and thresholds were invented, and no false-activation/hour rate is inferred. It demonstrates why a potential neural switch must be evaluated for across-session drift and continuous idle false-trigger exposure rather than only epoch-level classification.

The new artifacts were executed in a separate offline Python workspace: 13/13 isolated tests passed, with byte-for-byte Git blob verification against the branch. The full YazSes suite, Ruff and real-device checks have not been run. No production code or approved ADR changed. The next gate remains a license-reviewed, session/participant-separated public neural EEG dataset protocol and causal replay (#696 is closed, so this follow-up needs a maintainer decision to reopen it or to open a new issue; evaluation metrics sit under #817), with any actual human study separately gated.