Duplicate target columns reduced test Pz error at ridge penalty 100 in 68 ds006018 adults

We tested whether known block counts improve later Pz voltage estimates in a visual oddball task. The source was OpenNeuro ds006018 version 1.2.2. Each person's model used blocks 1 through 3 for training and blocks 4 and 5 for testing. The fixed primary method retained 68 adults and used ridge penalty 10. Two fixed-count information columns increased mean test error by 1.41%. The 95% gain interval was -3.47% to 0.37%. Thus the primary test showed no clear gain beyond stimulus identity and history. A planned penalty 100 test gave a 1.28% gain. A later exploratory control added two columns that carry no new target information. This control reduced reference error by 1.58% at penalty 100. Its descriptive 95% interval was 0.86% to 2.37%. Fixed-count features showed no clear gain over that control. We proved that three signed copies of the standardized target column reduce its effective ridge penalty by a factor of three. An independent solver confirmed this identity for all accepted test trials. The control needs an independent test for population claims. The package supplies a public EEG benchmark with source code, unchanged parent trial scores, all predictions, bootstrap indices, and pinned raw retrieval recipes. Independent checks reconstructed 19,200 source trials and verified 1,886 fits. Clean replays reproduced 97 numerical files exactly. Fresh EEG processing of two records reproduced all trial values exactly. The result applies to offline Pz scores in later blocks of the same people.

EEG · P300 · visual oddball · ridge regression · fixed-count information · duplicate predictors · prediction error · OpenNeuro ds006018