GRAPHOS re-examined: graph-theoretic features do not improve dyslexia screening, and a demographic confound in the Dytective dataset.

Introduction: Most machine learning methods for dyslexia screening treat the performance measurements of cognitive tasks as independent feature vectors. It has been hypothesised that this discards information carried by the interdependence among tasks probing shared cognitive processes. This paper tests that hypothesis and does not find support for it. Methods: We evaluate GRAPHOS, a graph-theoretic framework that builds task correlation networks from gamified cognitive assessment data, on the Dytective dataset of 3,644 participants (392 dyslexic). Model, aggregation strategy and decision threshold are selected exclusively on training and validation partitions, and the test partition is read once. Results: On the threshold-free AUC, GRAPHOS does not outperform a strong single-classifier baseline (0.8881 vs. 0.8883; DeLong p = 0.989), and this null replicates across ten random splits. A model using no graph-derived features is statistically indistinguishable from the full framework. A size-matched permutation null shows that none of nine population-level network differences between the dyslexic and non-dyslexic groups survives correction for the 8.3:1 sampling asymmetry. A subgroup audit shows that dyslexia prevalence in this dataset differs 4.7-fold by native-language status, that four demographic columns alone reach AUC 0.724 against 0.744 for all 192 task metrics, and that including them yields a twelve-fold disparity in false-positive rate between language groups. Discussion: A psychometric analysis indicates why the graph representation gains nothing: the battery exhibits little common variance in this sample (KMO 0.497, Cronbach's α 0.171), and exploratory factor analysis recovers 13% of variance with 28 of 32 tasks loading on no factor. We report these as negative results together with the two controls that produced them, and suggest both as routine: the size-matched null wherever correlation networks are compared between groups of unequal size, and the subgroup audit wherever a screening classifier is trained on a sample whose label prevalence may covary with demographic composition.