Abstract

Recent advances in camera-based artificial intelligence have enabled reliable quantification of strabismus from monocular video without explicit calibration. However, temporal models remain opaque, complicating clinical auditing and forensic failure analysis, particularly when systems face noise-induced binocular inconsistency (NIBI) caused by blinks, saccadic movements, or transient landmark-tracking errors. This study presents a forensic explainability audit of a validated calibration-free framework for automated prism diopter estimation. Rather than focusing on architectural modifications or accuracy improvements, the investigation targets the specific roles of temporal attention and stability-aware mechanisms in diagnostic decision-making. Temporal behavior is analyzed across multiple dimensions: quantitative attention attribution, alignment with an independent ocular stability metric, and inference-time perturbation experiments—including frame shuffling and contiguous occlusion—to assess reliance on sustained physiological evidence. Our findings show that the attention mechanism systematically prioritizes stable, fixation-like intervals while actively down-weighting transient artifacts. Prediction stability degrades when informative temporal windows are disrupted, reinforcing a reliability-weighted-aggregation interpretation in which the model values signal quality rather than strict sequential ordering. By providing a transparent and reproducible account of how temporal evidence is integrated, this study strengthens the interpretability of automated strabismus measurement, supports safer telemedicine deployment, and introduces a generalizable methodology for auditing temporal AI in clinical settings.