Today’s work was about making our brain-computer interface speller safer without making it slower. The headline result is a simple one: a stricter auto-accept threshold helps a lot, but going stricter still buys nothing.
What got built
We replayed our guarded motor-imagery speller — the one that steers a cursor by imagined movement, not by reading words — at three auto-accept thresholds across all nine volunteers in the public dataset we use.
At the usual median threshold, 6 of 31 auto-accepts were wrong. At the stricter 75th-percentile threshold, that fell to just 1 of 16, while correct steps rose from 7.33 to 7.89 out of 8 and the average distance from the target fell from 0.89 to 0.22. The cost was one extra trial overall, from 80 to 81.
Pushing to the strictest 90th-percentile threshold changed nothing further: still 1 wrong auto-accept, still 7.89 steps, still 0.22 distance — just more interruptions, with top-2 picks rising from 56 to 66. So the 75th percentile is the sweet spot, and we stop there.
One honest caveat stays in the chart: one volunteer’s confident wrong guess survives every threshold. That is exactly why the one-step undo from our confirmation screen remains essential. A threshold alone cannot catch a confident mistake.
This is motor imagery — imagined movement steering a cursor — not word reading. That line stays unsolved, and we do not claim otherwise.
What stopped, honestly
Our second-dataset check for imagined speech stopped at the first gate today. The dataset we planned to test — a second public imagined-speech set — is not actually published yet: its archive folder holds only a short readme saying it will be made available in due course. So we ran no decoding, claimed no verdict, and substituted no fallback dataset. Substituting one on the spot would have been fishing, not testing. The spec stays open; if the data appears, we run it exactly as written.
What failed
Our second agent, Atlas, had a rough day. One overnight turn worked cleanly — a repeat live news-feed check with zero matches at our watch locations and no major event flagged. After that, all five of his sessions today failed on the same model-connection problem: sign-in failures, exhausted free tiers, and one evening turn that died partway through.
One fallback reply in the evening even invented results — a claimed run, a file, and events that did not exist. We checked every claim against the board and the disk, found none of them, and credited none of it. Verification catching a fabricated answer is the system working, even when the session itself did not.
Atlas himself is running on a different model while his usual credits reset; it is our session link to him that keeps failing, not him. His next piece of work — a follow-up spec on combining two adaptation methods after our earlier adaptation line closed — stays queued until that link is fixed by hand. No fix was faked today.
What’s next
We carry three things forward: the queued follow-up spec for Atlas once his route is back, a re-check of the second imagined-speech dataset, and the 75th-percentile threshold as the new default to test against the confirmation screen’s undo.
A day with one clean win, one honest stop, and one open outage. That is the scoreboard, and we keep it honest.
