Abstract
Digital mathematics platforms support probabilistic prediction of subsequent student responses, but temporal ordering, score comparability and calibration affect the validity of evaluation. We evaluated AdaptiveMath-AI on a deterministic cohort of 385,037 Eedi interactions from 2904 students. Early representations were frozen before supervised fitting; learner states were reconstructed in recorded completion order, with simultaneous events predicted before batch updates. Twenty-two numerical inputs supported a gradient-boosting anchor, regularized item corrections, valid curriculum fallback and separately selected calibration. Six monthly expanding-window evaluations, repeated across ten seeds, covered 143,565 responses. The primary estimand averaged within-month performance differences rather than ranking scores from different fitted windows together. Mean monthly ROC-AUC was 0.761173 for AdaptiveMath-AI and 0.756833 for the 100-iteration comparator; the paired difference was 0.004339 (95% conditional student-cluster interval, 0.003636–0.005131). Mean monthly log loss and Brier score were 0.544721 and 0.184383. Independently calibrated ablations supported the contribution of question correction but not an additional hierarchy benefit. Against a recency-matched question-rate control, question-only residual correction improved ROC-AUC by 0.000628 (0.000304–0.000928). Calibration superiority was not uniform. The contribution is a temporally explicit, reproducible evaluation of item adaptation, not a new foundational architecture. Missing presentation timestamps and the previously studied single-platform source limit prospective, external and educational-utility claims.