In 12 SEEG subjects, the top 20% of working-memory trials by error score had 1.51× base error rate at delay end; only 48% were errors

A closed-loop memory aid must select the trials on which it acts. We measured the precision of a SEEG error decoder when it flags only the top k% of trials. We used the archived band-power features of 08fc8e7b (OpenNeuro ds006136, load-3 object working memory, 13 sessions, 12 subjects). A per-channel shrinkage-LDA ensemble gave out-of-fold error scores (5-fold CV, 5 seeds). Error = fewer than 3 of 3 items correct. Pre-registered primary test: at a 20% flag rate with encoding plus full-delay features (t = 6.5 s), the median lift (precision / base error rate) was 1.51. 11 of 12 subjects had lift > 1 (exact one-sided Wilcoxon p = 0.0005). Median precision was 0.48 at a median base error rate of 0.37. Thus, approximately half of the flagged trials were correct trials. Lift was 1.63 at 10% and 1.26 at 50%. The causal decoder at t = 3.5 s (1 s into the delay) was weaker: lift 1.36 at 20%, p = 0.065. Only 5 of 13 sessions passed a per-session permutation test. The analysis is offline and within-session, and the threshold uses the ranked held-out scores. A real system must set the threshold from past trials only. Package: numpy-only code, plan, per-subject tables, SVG figure, input checksums.

SEEG · working memory · error decoding · precision · lift · closed-loop · cross-validation · band power