Three sessions, read

Neurosity Crown-2B1 · 8 channels · 256 Hz · 248,320 samples · 16.2 minutes of signal

Three recordings came out of the console as raw CSV and were read from the samples up: a 1 to 45 Hz zero-phase Butterworth filter, then Welch power spectra in 4-second windows. Every figure on the three session pages traces back to a file, and the filter and spectral method were checked against a SciPy reference implementation before any of it was written down. That check is about this pipeline agreeing with a standard library; it is a different check from the one on the console-decoded page, where an independent reimplementation reproduced the console’s own signal-quality figures to within 0.1%.

These three were not planned as a series. They are the three recordings I had, and reading them one after another is what turned three ordinary session files into four rules I now apply to every recording.

The three

• Session 01, meditation focus to calm: the meditation that peaked at minute three. Five minutes, deliberately moving from focus into calm. The only one of the three where the filename was a hypothesis.
• Session 02, Seagreen donkey: ten minutes of eyes closed, and a state I kept sliding out of. The longest recording and the cleanest posterior contact. Alpha rose as expected, and so did something else.
• Session 03, Limegreen horse: the recording that only measured the headset. One minute, one dead electrode, and a spectrum full of movement. The session that produced no reading about me at all.

What the three have in common

This is the part that generalises. Nothing below is specific to a state, a day or a headset fit, and each one changed how I read the next recording.

The first minute is never real. Posterior alpha opens at 4.2% in the meditation and 3.2% in the eyes-closed session, then climbs steeply from there. That is me settling, not a brain state. Any average taken across a whole recording is diluted by it, so discard the first sixty seconds before comparing anything.

Delta always dominates, and that is not a finding about me. Even after filtering, 72 to 95% of the power in every channel sits below 4 Hz. Raw scalp EEG is like this: slow drift, breathing, tiny head movements and electrode settling all live down there. It means band shares are only comparable between recordings, never read as absolutes.

Amplitude is the fastest quality tell there is. A channel whose RMS is an order of magnitude above its neighbours is not a more active brain region, it is a worse electrode contact. That single check catches F5 in both Limegreen horse at 26 times the median and Seagreen donkey at 8 times, before any interpretation gets built on top of it.

A large DC offset is not a fault. Every electrode on this device sits tens of thousands of units away from zero and many never cross it, including the cleanest channel in the whole set. The high-pass removes it. Judge a channel by how much it moves, not by where it sits.

How to read the evidence on each page

Every session page carries the same four things in the same order: the numbers the file reports, the reading I drew from it, the electrode quality that licenses that reading, and the graphs the reading rests on. Session 01 carries one graph where the other two carry two, because its whole-session band table was withdrawn when an independent reimplementation of the pipeline could not reproduce it.

What the five bands mean, and which electrode sits where, is on Console Decoded rather than repeated on each page. The session pages assume you can look up alpha; what they add is what alpha did on one particular afternoon and whether the contact was good enough to believe it.

Where the trials went

These three recordings are the readable half of what the Crown produced. The other half is 219 training trials that cannot be opened at all. The log of how those trials were recorded is readable even when the trials are not, and it is written up separately in what the training log shows.

Sources

My own data: three recorded console sessions (248,320 samples), read from the raw CSV exports and computed with the filter and spectral method described above.