DOES SCRIPT CONDITIONING HELP DYSLEXIA DETECTION? A CAPACITY-MATCHED ABLATION ON URDU–ENGLISH HANDWRITING

Abstract

Handwriting images are an unusually accessible signal for early dyslexia screening, and the recent release of a bilingual Urdu–English children's handwriting corpus makes it possible to ask whether a model should be told which script it is looking at. We test that question and answer it in the negative, and we report first what we had to establish before the question could be asked at all. A content-level audit of the released corpus finds that its 852 files hold only 638 distinct images: one class folder contains 194 redundant copies, 20 images are byte-identical across both diagnostic folders and therefore carry contradictory labels, and the corpus's stated one-to-one class balance holds only at the file level. Deduplicated and with the contradictions removed, the corpus is 406 dyslexic against 212 non-dyslexic images, and the majority-class baseline moves from 0.500 to 0.657. On that corrected corpus we run a pre-registered, capacity-matched ablation: a script-conditioned channel-attention module against an identical-size control whose gates receive a learned constant instead of the script, under one fixed split, 30 shared seeds, two co-primary endpoints and a freeze tag applied before any test evaluation. Neither endpoint is rejected. The pooled accuracy difference is +0.0075 with a 95% confidence interval of [−0.0174, +0.0325], and the Urdu-subset difference is +0.0008 with an interval of [−0.0216, +0.0232]. We report the achieved power of this design honestly: it is 0.121 for the pre-specified effect of 1.5 accuracy points, so the correct reading is that the design could not detect an effect of that size, not that no effect exists. A counterfactual probe of the attention gates supplies a mechanism for the null rather than leaving it unexplained: holding each image fixed and varying only the script identifier moves the gates about 3% as much as changing the image does. We claim no novelty for script-aware computation as a mechanism. The contribution is the audit, the controlled test, and the mechanism, in that order.