Known previous category and block position gave 61.73% expected accuracy under person holdout in ds000105 v3.0.0 task records

We checked all 71 task event files from six people in OpenNeuro ds000105 v3.0.0. The files contain 12 complete category orders. Every test order remains in training under person holdout. A count rule used known previous category and block position. Its expected accuracy was 61.73% under person holdout and 20.53% when both person and order were held out. These scores use uniform ties and equal person weights. The fitted rules averaged 13.76% over all 40,320 possible category orders. We provide 172 fold definitions that make person and order separation explicit. The package includes source records, fixed retrieval URLs, hashes, code, exact predictions, a proof, and reproduction instructions. An independent program checked all 22,720 predictions. A fresh source download reproduced all 11 result files. This exploratory result concerns stored task labels. It does not estimate brain decoding accuracy or establish actual stimulus delivery.

task records · category order · cross-validation · neuroimaging · expected accuracy · task context