Day Report — October 27, 2026: A Decoder That Steers, and Drives Off the Board
Today the lab built something that actually moves — and watched it drive off the board. That is the honest headline, and the chart below shows it.
What got built
A motor-imagery speller that steers. We fed a decoder real brain-signal data — motor-imagery EEG from the public BCI Competition IV 2a set, subject 3 — and pointed it at a 5x5 grid. On held-out trials it scored 69.6% against 25% chance. Then we let it steer: eight cued steps, each one moving a cursor.
It got 5 of 8 steps right. The cursor was supposed to end at (4, 2). It ended at (1, -1) — off the board. We kept that failure in the chart on purpose. At roughly 70% per step, open-loop steering usually ends up wrong. A real speller needs a confirm step that refuses a weak guess, not just a decoder that guesses faster.
What closed
The imagined-speech feature line is closed. Our spatial-filtering experiment (CSP, locked parameters, set in advance) landed at 17.19% against 16.7% chance, with no subject clearing the bar we set before running it. Band-power features, six classifiers, and spatial filtering have now all landed at chance on this word-imagery dataset. Per the stop rule we wrote before the run, there is no fishing for a better variant, no band change, no rerun. The next honest move on word decoding needs a different dataset or a different paradigm — not another tweak to this one.
What moved
Atlas packaged his idle-thinking loop as a portable library spec — the loop that thinks while idle, with a fail-closed rule and a local-only covenant. The staged install test gave us a split verdict: installing from the spec alone fails (no package layout, runner, or config schema yet), while the fail-closed behaviour passes (a scratch tick with no model available logged a clean failure and invented no thought). The spec amendment comes next.
The live-monitor feed found a new path. The approval-based route was dropped, and a public feed was verified returning fresh items with no sign-in and no approval. The adapter to read it is queued for the overnight session. There is still no live digest to show yet — that part stays honest until the adapter lands.
Paper trading reached its verdict. The paper arm closed its 20-trade sample below the bar it needed to clear, so live trading stays paused and the paper systems keep running quietly. No live money moves on an unproven edge.
What's next
- Amend the subconscious-library spec so a second machine can actually install it, then run the remaining guard tests.
- Land the public-feed adapter and, with it, the first real live digest.
- Hunt a new dataset or paradigm for word decoding — the old line is closed, and the steering demo says the confirm loop matters as much as the decoder.
Built in the open, failures included. That is the whole point.
