0.2.0¶

Signal-processing accuracy & honesty¶

• Per-reading confidence. Every BrainState now carries confidence (0..1), per-metric metric_confidence, and a status (ok/warming_up/unreliable), folding in signal quality, calibration, and window fill. Surfaced through the MCP get_brain_state, get_signal_quality, and stream_summary tools.
• Artifact handling, not just detection. Hard artifacts (flatline, railing) or poor signal mark a reading unreliable and collapse its confidence, so contaminated windows are never narrated as clean. Added an emg artifact flag for implausibly large amplitudes.
• Re-grounded metrics. focus is now documented as the genuine Pope et al. (1995) engagement index β/(α+θ); engagement is a distinct β/α arousal ratio (no longer mislabeled "Pope-style"); fatigue uses the Eoh et al. (2005) (θ+α)/β index; meditation is relative alpha α/(α+β+θ) — decoupled from theta so it no longer confounds with drowsiness/fatigue. Delta and gamma are excluded from metrics (drift/EMG-dominated on consumer hardware). Each metric ships a formula + basis + caveat via the new get_metric_definitions MCP tool.
• Calibration fix. Bounded metrics (calm, meditation) are centered at 0.5 uncalibrated instead of being capped low; unbounded ratios stay centered at 1.0.
• DSP fixes. Longer default analysis window (2 s) with Welch segment averaging for stable low-frequency band power; narrow-band single-bin integration no longer silently returns 0; notch is applied before the bandpass.

0.1.0¶

• Device-agnostic EEG acquisition: synthetic, NeuroFocus (serial + BLE), OpenBCI/Muse via BrainFlow, LSL, generic serial, recording playback.
• DSP pipeline: band powers (δ θ α β γ), focus/calm/attention/engagement/fatigue/meditation metrics, signal quality + artifacts, calibration.
• Real Model Context Protocol server (FastMCP, stdio) with tools, resources, and a prompt.
• CLI: devices, stream, record, play, neurofeedback, dashboard, serve.
• Recording (CSV/npz/EDF) + replay, neurofeedback trainer, LSL publisher, FastAPI web dashboard.