Benchmarking Observable Platform Dynamics Beyond Persistence in Three Open Educational Datasets

Educational theories of artificial intelligence (AI)-supported learning include mechanisms absent from reusable platform logs. We separated a conceptual framework from four source-derived proxies: cumulative achievement, engagement, resource breadth, and digital activity. The Open University Learning Analytics Dataset (OULAD) was primary; Eedi provided external evaluation and EdNet an exploratory transport stress test, with 72,277, 7149, and 513 lagged transitions. In the OULAD temporal holdout, observable-state ridge regression reduced root-mean-square error (RMSE) versus persistence by 14.51–15.59% for behavioral proxies but by 0.16% for achievement. Autoregression captured 98.24–99.78% of behavioral gains; cross-state predictors added 0.038–0.324% beyond autoregression. Three outcomes supported prediction beyond persistence; none supported incremental cross-state prediction or coefficient interpretation. Same-feature gradient boosting improved behavioral-proxy prediction by 0.84–2.84% but worsened achievement by 3.47%. Eedi resource breadth improved by 21.38% versus persistence and 1.71% versus autoregression; EdNet resource breadth deteriorated by 152.96% and 15.32%, respectively, among 82 learners in one group. Performance was outcome- and dataset-specific. The study did not test generative-AI effects or causal educational mechanisms; formal measurement invariance was neither assessed nor established. Longitudinal platform indicators can inform pedagogical inquiry only after local validation of construct meaning and calibration.