Changelog

Format: Keep a Changelog. Changes to the flashed sketch are listed under Board so a recording can be matched to the firmware it was made with.

[Unreleased]

No model-v* tag: quality.py is in the measuring chain, but spectrum() feeds only the workbench's drawing (POST /spectrum). Nothing it returns is stored in a session or an export, and the signal check (assess(), bands()) is unchanged, so no experiment is invalidated.

Fixed

• quality.spectrum() evaluated one DFT bin per display bucket and skipped the rest, so on a record longer than a couple of seconds a narrow line fell between the evaluated bins and vanished from the Monitor tab's spectrum (a 260-count 10 Hz sine over 20 s drew as about 1 count). Each bucket now reports the maximum amplitude of every bin it covers, from one zero-padded FFT of the most recent 16 384 samples at most. freqs are now bucket centres. resolution_hz keeps its meaning (rate / samples used); bucket_hz, bin_hz, samples_used and summary are new.

[0.2.0] — 2026-09-25

First public release. Tagged model-v0.2.0: this version changes numbers in recordings and exports, so results from 0.1.0 are not comparable without saying so. No preregistered experiments existed before it, so none is invalidated.

Changed (measuring chain)

• The signal check is readout-aware and called a signal-plausibility heuristic, not a contact or impedance check. ECG and EMG are no longer judged by EEG rules (alpha SNR, the 100 µV floor). Band estimates use a Hann window, with non-overlapping band edges.
• µV per count (7.9) is labelled with its real provenance: the shield's built-in cal signal at nominal amplitude, not externally checked. The circular "independently recovered" claim is gone.
• BIDS: PowerLineFrequency from the session, not a hard-coded 60 Hz; ECG/EMG channels typed correctly; ScalingProvenance, StreamIntegrity, SoftwareVersions; a root README; synthetic and real sessions never share a dataset root.
• The bridge parses the board's t_us column, counts malformed lines and timestamp gaps (handling the micros() wrap), and writes t_us into its own recording.
• The EEG montage puts the active electrode at Oz (not the inion). The EMG montage is a SENIAM bipolar pair on the muscle belly.

Added

• apps/server/experimental.py: every experimental route, marked "experimental": true, and "beta — experimental" badges in the workbench. See docs/experimental.md.
• KIT Adaptive Preregistration (core, prereg), EXPERIMENTS.md and docs/versioning.md.
• Sessions record software_version and mains_hz.
• scripts/check_docs.py: the README's safety list matches the workbench, and the blog post's shared figures and photos match this repo.

Fixed

• The bridge bound to all interfaces and answered any web origin by default. It now binds 127.0.0.1 and answers only loopback origins (or $OLIMEX_ALLOWED_ORIGINS).
• The signal panel's readout row never rendered. The rate row now says declared and shows the measured rate separately.
• UI text no longer attributes the noise spread to electrode prep (it is unexplained), or calls a 40 Hz ASSR "within reach" (it is not yet shown).

Safety

• The list now covers unplugging every mains-connected device, isolating the USB data line when not on battery, not touching grounded metal, and any current path across the chest.

Board

• eeg_stream.ino: header comments only; the flashed behaviour is unchanged.

Housekeeping

• ruff lint and format at line length 100, enforced in CI. References to the private repos this code came from are removed.

Moved out (to optional companion tools, found via $OLIMEX_TOOLS)

• The Auditory Response Atlas, the extraction layer (lock-in, noise budget), injection curves and the acquisition-requirements checks. The workbench shows what depends on them only when they are present. This board's noise figures stayed here.
• The built-in S3 uploader was removed; storage.py is a plugin hand-off only.
• quality.py carries its own Goertzel helper so the contact check never depends on an optional tool.

Fixed

• A plugin uploader was handed the live session folder, so an earlier upload's uploads.json receipt went into the snapshot. It now gets a staged copy holding exactly the manifest's files.

Verified

• Live upload of a synthetic session to S3-compatible storage through a plugin. Every object's ACL was private, and every object refused a request made with no credentials (HTTP 403). The test objects were deleted afterwards.

Changed

• Copyright holder is Polarize.Tech (LICENSE, CITATION.cff, pyproject.toml).

[0.1.0] — 2026-09-15

First release as its own repository. The code was extracted from earlier internal tools.

Added

• apps/web: the workbench. Live view with the Atlas tab, recording and tagging, playback against the tag track, electrode placement guide, uploads, BIDS export.
• apps/server/storage.py: private uploads to S3-compatible storage (DigitalOcean Spaces, AWS, MinIO) or a plugin, checked from outside with an anonymous read.
• apps/server/bids.py: stdlib EEG-BIDS export (BrainVision), verified by reading it back with MNE-BIDS.
• apps/server/timebase.py: the bridge measures the board's real sample rate against the host clock.
• sessions.annotate(): adds the measured rate to a finished session without rewriting it.

Fixed

• The recording panel dropped the last partial batch at Stop: 1859 samples stored against 2020 in the black box on the real board. Samples buffered while the store opened were also lost, and appends could race each other. Now contiguous and ordered, verified as an exact slice of the black box.
• Every recording's black box was written to capture.csv, overwriting the previous one. Each now gets its own name.

Board

• eeg_stream.ino is unchanged from the version checked against the live board on 2026-09-14.