Open architecture is the foundation for verifiable specifications
A specification you cannot audit is a claim. What has to be open for an EEG device to be checkable — and what we publish so ours is.
November 2026BROADMIND
Every EEG device arrives with numbers: a noise floor, a sample rate, a channel count. Those numbers are usually true in the sense that the vendor measured something, and almost never true in the sense that you could measure it too. The gap is not dishonesty. It is architecture.
The chain nobody can see
Between a microvolt at the electrode and a float in your file there is a chain: amplifier gain, reference selection, filtering, packetisation, and whatever the firmware does when its buffer fills. On an open device every step is readable. On a closed one the chain is a compiled binary — the single artefact that decides the scale, the filters, and what happened to the samples that never arrived, with no read access.
That is the difference between a specification and a claim.
Four questions a closed device cannot answer
What is the scale? Counts become microvolts through reference voltage, gain and the converter's full-scale convention — all of it in firmware. You accept a factor and hope it matches the binary that was actually flashed.
Is the rate real? "250 samples per second" is a setting, not an observation. Whether 250 arrived needs a counter inside the data.
What was lost? A dropped packet and a clean stream look identical on a link with no sequence number. You find out months later, if at all.
What was filtered? An undisclosed high-pass is the quietest way to change a result. A 0.5 Hz cutoff removes slow cortical potentials and drift alike.
Open has to mean four things
Publishing source code is not enough on its own. For a claim to be checkable, four things have to be open: the protocol, versioned, so bytes can be matched to samples; the decoding path, so it can be reimplemented; the test vectors, so two implementations can be compared; and the provenance, so a file states what produced it.
What we publish
The protocol. A documented, versioned service definition with an eight-byte header carrying a sequence number and a CRC-16, so loss and corruption are reported rather than hidden.
The SDK. Open source, MIT, with a replay transport so a decoder can be tested without hardware.
Test vectors. Raw notification bytes in, expected samples out — the same vectors our own tests run against.
Provenance in every export. Gain, reference, filters applied, firmware version, protocol version.
Two exits. LSL and BrainFlow, so data can leave our software entirely.
The safety limit beside the specifications. Battery-only operation, no galvanic isolation.
What we have not published yet
The protocol above is a specification, not shipped firmware. The wireless bridge does not exist yet, and no raw capture from a production unit is committed to the repository. We are stating the standard before meeting it because it is more useful as a commitment than as a retrospective.
The test
A device should answer three questions without a support ticket: what is the scale, and how do you know; what happened to the samples that did not arrive; and what was done to the signal before you see it.
A .hex file tells you what the vendor decided. A protocol, a test vector and a provenance block tell you what the device does.