Three sessions, read · the training log
What the training log shows
The 219 trials cannot be opened. The log of how they were recorded can, and it explains more than the trials would have.
17 experiments · 219 trials · 7 of them holding anything
Three systems, three shapes, and none of them talk
Before any numbers, the map, because the thing that makes this data feel unreadable is that it is split across three systems with incompatible shapes.
Studio holds the labelled trials. Seventeen experiments, of which seven contain anything: 219 trials across push, leftHandPinch, artifactDetector, mentalMath, drop, jumpingJacks and bitingALemon. The console shows a count and nothing else, with no trial list, no per-trial scores and no export button. Those 219 trials exist and cannot currently be opened.
Recordings holds the raw signal. Twenty recordings, about an hour in total, exporting as CSV, JSON or EDF. This is the readable data, and it is what the three session studies read.
The crown-debrief project expects a third shape entirely: one row every two seconds with named focus and calm columns. Its parser rejects any file without a focus column, which is every file the console produces. Every row in its data/ folder is marked mode=mock, and its own handoff notes say no real headset data has passed through it.
The gap: The bridge between the second and third of these is the missing piece. Reading a console export and turning it into the same kind of plain-language debrief, without needing the schema the project cannot get, is the work that closes it.
Every experiment, in order
The 219 trials cannot be opened: no export, no per-trial view, nothing held in the page, nothing the MCP server will return. But the log of how they were recorded is fully visible, and it turns out to explain more than the trials would have. Experiments in the order they were created. Bars are recorded trials. Hollow marks are experiments that were created, sometimes with labels chosen, but never recorded a single trial.
Not one multi-label experiment ever recorded a trial
This is the finding that matters most. All 219 trials sit in single-label experiments: 30 of push, 30 of leftHandPinch, 30 of drop, 30 of jumpingJacks, 31 of bitingALemon, 60 of mentalMath, 8 of artifactDetector. Every experiment where I selected two or three labels together is empty.
A single-label experiment teaches the system one thing: this command versus rest. That is a useful detector, and it is what Kinesis is built around. But telling two commands apart, the thing you need for an interface with more than one action, requires both classes recorded in the same session, under the same cap placement, in the same sitting. Those are exactly the three experiments I set up and never ran.
The three abandoned multi-label sessions
| Labels chosen | Why this pairing would have struggled |
|---|---|
| tongue · leftHandPinch · jumpingJacks | Tongue imagery sits low and lateral, nearest T7 and T8, which this headset uses as reference and bias rather than as recording sites. |
| leftHandPinch · leftArm | Both are left-side movements, so both project to the same right hemisphere. Nearly the hardest pair to separate. |
| rightHandPinch · rightFoot · rightArm | All right-side, so again one hemisphere. Foot imagery also localises to the midline, and there is no midline electrode on a Crown. |
Why those pairings were the hard ones, and what to record instead, is the argument in command selection as a constraint problem.
Two of the chosen labels were never going to work on this hardware
The Crown records at CP3, C3, F5, PO3, PO4, F6, C4 and CP4, with T7 and T8 spent on reference and bias. There is no midline sensor. Foot movement is represented on the medial wall of the motor cortex, directly under the midline, so rightFoot is the one label in the library this montage is least able to see. Tongue is represented low and laterally, closest to the two positions the headset does not record from. Both appear in my abandoned sessions, and neither was a fair test of my ability to do motor imagery. The anatomy behind that is in two commands that you can't train.
Training stopped, and the log shows how
The last trial I ever recorded was in the ninth experiment. Eight more have been created since, and every one of them is empty.
The last cluster is the clearest. I created three experiments inside eight minutes, and then, nine seconds after the third one, started the five-minute Focus to Calm recording instead. Three attempts to begin training, then a recording. Whatever stopped me was in the way of starting a session, not in my willingness to sit down and do one.
What the 219 trials are actually worth
They are not a dataset I own. They trained models inside Neurosity's cloud that I can only reach through live prediction, never inspect, re-analyse, or take elsewhere. About 90 are imagined movement, though only the 30 hand pinches land where C3 and C4 can read them; push and drop load the frontal planning regions instead. Sixty are mental arithmetic, which is a cognitive-load task rather than an intent. Thirty-one are bitingALemon, which is largely jaw muscle, as I noticed at the time.
Their real value is as a record of what I tried and what went wrong, which is what this page has been able to read. Everything else needs labelled trials recorded and kept by whoever records them, which is the missing collector in the crown-debrief project rather than a second thing to build.
Sources
My own data: an audit of seventeen training experiments in the Neurosity console, read from the training log, and the crown-debrief repository.
Explore more
This site covers what the documentation doesn't: the things I wish someone had handed me first.