Encoding SEEG power-pattern similarity does not predict the next free recall when lag and position are controlled

We tested a prediction of temporal context models: if encoding brain states drift and recall reinstates them, similarity between encoding patterns should predict which word a patient recalls next. We used word-locked SEEG log power (5 bands, bipolar pairs, 0–1.6 s after word onset) and recall order from ds004789 events.tsv, with SHA-256 checks. For each transition between recalls, we calculated the percentile rank of the similarity to the actual next word among words not yet recalled but recalled later. We removed serial-position means, subtracted mean similarity for each lag, and compared results with position-matched surrogate lists. The analysis plan was written before neural results. Of 54 subjects, 31 had at least 20 transitions (1,376 transitions). Behavior showed temporal contiguity (temporal factor 0.676, 95% CI 0.637 to 0.718; 28 of 31 above 0.5). Encoding similarity decreased with lag (slope −0.0022 r per lag step, p = 1e−5; 27 of 31 negative). The primary effect was −0.015 rank units (95% CI −0.038 to 0.007; sign-flip p = 0.20; 14 of 31 positive). Raw similarity (0.509, p = 0.53), lag-residual similarity (0.491, p = 0.52), and theta-only and HFA-only analyses also showed no effect (Holm p ≥ 0.73). In these band-power features, measured drift did not organize recall order. The study covers session 0 and clinical electrode coverage; it does not test time-resolved patterns or recall-period reinstatement, and small effects (<0.01 rank units) remain possible.

free recall · SEEG · temporal contiguity · episodic memory · band power · context models · serial lag · neural drift