EEG/EOG activation feasibility: evidence review and decision gates¶

Date: 2026-11-03
Status: Research synthesis for review; no human/device study conducted in this review, no accepted new runtime decision.
Issue: #696, under #692.
Companion: Existing protocol; proposed ADR-v2-154.

Summary and recommendation¶

Continue narrowly as offline, non-actuating feasibility research; do not add a consumer-EEG activation source to YazSes. The current repository proves that a hand-authored synthetic signal generator and detector can be replayed deterministically and scored against injected events. It does not provide evidence of human EEG intentional-control performance, comparative advantage over EOG/EMG, or tolerable false activations in everyday dictation. Proposed ADR-v2-154 records the corresponding review gates; it does not supersede accepted ADR-v2-129 or the ADR-030 feature freeze.

A first useful research result would be a falsification: can the apparent EEG switch signal be explained equally well by EOG or EMG? Until this is tested with independently recorded channels, classify the output as an electrode-derived candidate with unknown physiological origin, never a decoded intention. This distinction matters particularly when YazSes users speak: jaw and facial muscle activity may confound scalp channels.

1. Repository baseline: inspected on 2026-11-03¶

Precise reconciliation of current protocol: Its P0 wording calls the source “BrainFlow's synthetic board and generated signals”, but the #696 replay currently uses Python's seeded random generator only. BrainFlow officially provides a distinct synthetic hardware-free board (board id -1) and a playback-file board, which could be independently smoke-tested later. Do not report a BrainFlow integration test as completed because the Python generator is green. BrainFlow's synthetic board is a development signal source; deterministic, byte-identical runs require either an independently seeded fixture or verified recording/playback, not an assumption about the live synthetic stream.

Other concrete gaps (not defects proven in production): - The replay emits activation times and score aggregates, not the proposed full timestamp / quality / confidence / age / abstention candidate contract. Do not confuse this with the canonical #699 multimodal schema. - The protocol asks for median and p95 latency, but the script currently reports median and maximum for matched synthetic events, not p95; no timings from physical devices are measured. - The generator's noise/leakage amplitudes and event spacing are chosen parameters; it has no genuine electrode movement, neurological variations, between-day drift, real speaking, or real hardware clocks. - No head-motion or quiet-reading test block is present in the #696 replay's fixed block list despite being identified as desired controls by the protocol. - There is no preregistered real-world pass/fail threshold, no held-out person or session comparison, and no evidence of independently observed EEG-only benefit.

2. Primary-source research findings (checked 2026-11-03)¶

Source-status note: These are documentation/library and governance sources, not a performed literature meta-analysis or head-to-head empirical review. Any quantitative third-party EEG claims must be checked against the original paper's population, equipment, task, confidence intervals and false-activation measurement before using them in YazSes product decisions.

3. Competing hypotheses, baseline and falsifiers¶

Explicitly separate neural EEG from ocular EOG and muscular EMG, including ambiguous mixtures. A detection on a scalp electrode is not by itself evidence of cerebral origin. Results should label the physical origin as unresolved unless the study design supports the distinction.

4. Repeatable work plan, without new runtime privileges¶

S0 — Static reproducibility and source separation (permission: ordinary offline engineering)¶

1. Run the existing #696 Python replay and tests without a sensor, record the repo SHA, Python version, seed, complete parameter table and output hash. Run multiple predeclared seeds; report counts, not inferred human accuracy. These commands are instructions for future reproduction, not a claim that this review executed them:

   uv run python scripts/hci_eegeog_replay.py --seed 0 uv run python -m pytest tests/test_hci_eegeog_replay.py -q

2. In a separate research-only script (not imported by the daemon), create BrainFlow BoardShim SYNTHETIC_BOARD and check device lifecycle, channel metadata, units, timestamp monotonicity and shutdown; test PLAYBACK_FILE_BOARD with an explicitly permitted local synthetic file. Pin BrainFlow 5.23.0 as the initial version to verify, not a newly added project dependency. Do not assert that BrainFlow's synthetic signal has a controllable seed.

3. Compare the #696 candidate-output shape with the actual #699 fixture and #139 intent vectors; write a compatibility mapping, not a new public contract.
4. Add missing adversarial fixtures: reading quietly, head motion, electrode pop, flatline/dropout, clock jump/out-of-order samples, stale source, loss of quality and simultaneous stop request. No OS calls.
5. Reconcile protocol-versus-script metrics (p95 vs maximum) and expose “unknown confidence/quality” rather than inventing calibrated probabilities.

S1 — Synthetic stress and provenance (permission: no sensor)¶

Freeze parameters before each batch; sweep signal-to-noise, EEG/EOG/EMG leakage, threshold, refractory, hold duration and drift independently. Use distinct seeds for calibration/development/evaluation and retain all failed cases. Report per-condition numerator/denominator, missed switches, wrong-origin attribution, abstentions, latency percentiles, CPU/RSS, and reproducibility metadata. Never convert the synthetic event schedule into “false activations per hour of real usage.” Include a default-off and no-actuation assertion.

S2 — Public/previously licensed human datasets (only after separate provenance review)¶

Optionally reproduce a MOABB-compatible baseline on a dataset with an audited license and consent scope, entirely in an isolated local analysis environment. No automatic downloads; acquisition must be explicit and logged. Predefine subject/session split, channel montage, preprocessing, causal vs offline transformations and leakage checks. This step tests analysis methods, not in-the-wild safety or permission for a consumer feature. Do not commit any participant recordings or derived sensitive time series.

P2 — Real-hardware shadow-mode study (NOT AUTHORIZED BY THIS REVIEW)¶

Proceed only after an independent ethics/review determination, named responsible investigator, accessible consent and withdrawal materials, risk/retention plan, hardware/electrical-safety assessment, and explicit owner approval under #425/#427 and ADR-v2-150. No flashing/SSVEP stimulation without its own photosensitivity review. Use a physical/independent stop, an observer stop, and a passive UI with no OS action injection.

Pre-register: - The exact intended activation task and population, inclusion/accessibility accommodations, hardware/firmware/montage, calibration time, exclusion criteria and whether data can leave a participant's device (default: no). - Separate idle, speaking, reading, natural blink, deliberate blink, clench/chew, head motion and recovery phases; log usable exposure after quality gating, plus excluded time and its reasons, so discarding difficult periods cannot make a detector appear safer. - An independent EOG and/or EMG reference when physically feasible, artifact-label adjudication rules and an EEG-only / EOG-only / EMG-only / combined ablation. - Frozen event-matching window, refractory, max stale age, quality thresholds, session/participant holdouts, handling of overlapping events and a rule that unknown origin means abstain. - Primary outcomes: accidental proposed activations per usable hour by phase (count, denominator and interval), cue hit/miss, p50/p95 end-to-end latency including acquisition, silent/paused failure response and confirmation/recovery burden. Secondary: setup time, fatigue/discomfort (voluntary report), CPU/RSS and accessibility. - One-sided upper uncertainty bound for unintended activations as a decision statistic, with an agreed false-activation budget established before recruitment, in addition to a lower uncertainty bound for hit rate. Under a stationary Poisson approximation, zero events over T hours gives an approximate 95% upper limit of 3/T events/hour; correlated bursts or drift require more conservative treatment. The budget remains undecided, and this formula is not a claim that current YazSes meets it.

Do not aggregate different hardware/person/session conditions into a single flattering mean. Report failures and dropout separately.

5. Threat and failure review¶

6. Decision ledger and next actions¶

Suggested handoff: review proposed ADR-v2-154; if accepted as a research policy, file narrowly scoped G1 work against #696 rather than reopening completed fixture work (#699); identify an eligible research review authority before considering P2. Do not close #696 based solely on this review; no new human/device evidence is supplied here.

References¶

• BrainFlow supported boards, User API, data-format metadata, releases, MIT license.
• MNE-Python: repairing artifacts with ICA.
• MOABB cross-session and cross-subject evaluations.
• OECD recommendation on responsible innovation in neurotechnology and NIH BRAIN neuroethics guiding principles.

All external sources were inspected 2026-11-03. This document contains no new empirical measurements.