VisionWeaver: CFS-Inspired Assessment and Adaptive Training Prototype Under DR-Inspired Simulated Vision Degradation

Abstract

VisionWeaver did not begin as a claim about diabetic-retinopathy diagnosis. It began with a smaller engineering problem: after a low-contrast visual-task session, a single score did not tell us enough to debug the run or explain why the software had made a difficulty change. We built a Python/PsychoPy prototype around that problem. It includes a side- by-side, CFS-inspired assessment task, a central-gate training task, keyboard fallback for development runs, and Arduino-compatible input when external buttons are attached. The first check was therefore about traceability. We used internal logs, simulated respon- ders, ablation, parameter sweeps, and failure-log review to ask whether the task could preserve a usable trial record. In Monte Carlo tests, staircase and moving-window con- trol reached 35.8% and 22.4% band occupancy; fixed difficulty reached 0.1%. A random forest using the first 10 synthetic trials classified simulated degradation level with 72.9% accuracy. Its warm-start setting raised band occupancy by +15.5%, reaching 98.7% of the Oracle baseline. These results support the software and analysis workflow, not clinical efficacy.