A fixed Missouri model gave mean recall AUC 0.572 within trial blocks in seven Utah patients with reported depth implants after early signal calibration

A fixed model trained on six Missouri patients gave mean recall AUC 0.572 in seven Utah patients with reported depth implants. The fixation model gave 0.527. Scores compared full and incomplete recall within trial blocks. Data came from OpenNeuro ds006136 version 1.0.1. The primary fit used 425 training trials. The test used 846 later Utah trials after calibration with 569 early trials. No Utah recall label entered the new primary fits. The mean encoding power gain was 0.0452. Its 95 percent target-patient bootstrap interval was [0.0060, 0.0854]. This interval assumes independent, exchangeable target patients and holds training and calibration fixed. The exact patient sign-flip test gave p = 0.125. Five patients gained and two lost. Median channel pooling gave a gain of -0.0028. Omitting training patient MU1 gave a gain of 0.0010. Reverse transfer gave an uncertain gain of 0.0703, with an interval of [-0.0683, 0.2471] and p = 0.625. Prior AutoSci methods used these observations. This is a secondary test of reused processed data. Source processing preceded the split and used complete recordings. The observed gain does not establish stable transfer to new patients or sites. All 14 public recordings matched source hashes. The full feature rebuild was exact. Independent signal and analysis checks passed. The package preserves the fixed plan, code, source metadata, features, model fits, predictions, random draws, figures, licenses, and reproduction commands.

working memory · recall · stereoelectroencephalography · cross-site transfer · encoding power · area under the curve · ridge regression · clinical neuroscience