Brain Speeddial Eeg Jev

No. Jev never sees EEG or decoder probabilities. It receives only the proposed action, its cost, and context such as time, location, and minutes since the same action, then answers whether the action makes sense now.

They are measured replays of every held-out trial across 10 subjects and 5 folds as continuous sessions. Confirmation cuts accidentals sharply but costs completion and adds seconds of latency, with repeat at 68% and imagery near 45%.

No headset: run_demo.sh replays public Shin 2017 recordings sample by sample, and the first run fetches datasets A and B through MOABB, about 5 GB, then replays subject 3 at 2x for roughly 40 seconds. Set TYPESAFE_API_KEY in .env for the live Jev gate; without it a labelled rule fallback decides fire, confirm, or ignore. The --metrics pass over 10 subjects x 5 folds x 3 confirm modes takes about 8 minutes.

Repeat-the-command is the practical default with higher completion, while left-right imagery is safer for money actions but loses more genuine commands because imagery decodes near 70% for most subjects.

Real: the EEG, which is public Shin 2017 A and B, 10 subjects and 30 channels, replayed one sample at a time; the mental task, since the subject really was subtracting; the decoders, trained on the other four folds and scored on held-out trials only; and the Jev policy gate, which makes live cached API calls. Simulated: context such as time, location and minutes since the last order. Mock: the agent, which writes text logs and orders nothing.

Because the offline experiments showed it is the wrong tool for that job: AUC 0.58 against 0.64 for a plain probability threshold at judging decoder evidence. It is useful at the other task, judging whether an action makes sense in context, where it reaches AUC 0.70. A separate EEG-to-descriptor-to-Jev experiment is left out of the demo because a plain classifier matches it.

One yes/no question, does this action make sense right now, answered from the proposed action, its cost and the context. With an identical order placed four minutes ago it returned p = 0.04 and blocked the action; with the default Friday-evening-at-home context it returned 0.71 and routed a money action to confirmation instead.