Neuroscience

A colleague asked for an honest assessment of training LLM/ML on an OpenBCI stack. I recommended small specialized decoders (band power / CSP / Riemannian / EEGNet) emitting a constrained state vector, with an LLM only narrating that vector — never trained on raw EEG. Real on Cyton: eyes open/closed, SSVEP, coarse alertness, P300 with a stim protocol, EMG. MI is subject-specific and borderline. Valence and thought-reading are not real.

OpenBCI GUI has no traveling-wave / Hilbert-phase / cross-correlation lag analysis; Head Plot is amplitude only. Cyton 8ch can do crude bilateral homologous-pair lag; 16ch daisy for crude direction; Ganglion 4ch cannot map travel.

Per-hemisphere laterality via band-power L vs R (laterality index on homologous sites; e.g. C3/C4 mu, O1/O2 alpha, F3/F4) is considered real and shippable on 16ch; treat as named-band change (often desynchronization), not broadband amplitude. Head Plot will not provide a laterality time series.

The standing session reference is an 8ch Cyton pre-start gate (250 Hz) that must pass every session before Focus or any decoder: SRB2 midline only (Cz, inion, or linked mastoids) with Bias on; 30–60 s contact (ignore GUI kΩ if untrusted); ~2 min functional check with eyes-closed occipital alpha if O1/O2 are in the map; mark rest/blinks/clench/brows/swallow/head turn without tuning on those; 1–40 Hz + 60 Hz notch only after green; 1 min open + 1 min closed as the session baseline (Focus is relati

OpenBCI GUI graphs use display-only smoothing; recorded BDF stays raw (1–40 Hz + 60 notch). Time Series default 10 s span, min-max or decimate draw to ~30–50 Hz, freeze µV scale after the contact gate (no per-frame autoscale). Band Power/FFT: 2 s window, 250 ms hop, Welch 2–4 segments, EMA τ ≈ 1 s, bars at 4 Hz. Focus: EMA τ 1–2 s, score ≤2 Hz. Head plot temporal smooth 5–10 Hz max. Raw unfiltered view remains a toggle for the pre-start gate.

Prefer EEG-only robot-arm control (no EMG). Extra channels do not buy depth (volume conduction); 8→16 helps scalp sampling/CSP, not deep sources. Path: SSVEP via FBCCA/TRCA as discrete shared-autonomy OSC commands (flicker widget), not continuous kinematics, MI first, sLORETA, or EEG foundation models. 8ch enough if O1/O2 are on the head; daisy later for MI. GUI is missing the stimulator.

EEG control vector is CCA/FBCCA (sine/cosine template correlation) or TRCA/CSP scores plus a pick, emitted as {target, conf} over OSC — not ODE/PDE integration of traces.

SSVEP flicker for an OpenBCI GUI must be real luminance locked to monitor refresh (on 60 Hz: 10, 12, 15, 20 Hz), not a soft opacity fade; 4–8 image targets; selection is gaze, not silent intent.

Use when starting an OpenBCI Cyton session (8ch default). Fail-closed contact and functional gate before Focus, record, or any decoder. 16ch laterality is the next step, not part of the go/no-go.