CHM-DTKT: a causal hypergraph selective state space digital twin for deconfounded and temporally grounded knowledge tracing

Adaptive tutoring is a complex intelligent system in which learner competence, exposure policy and memory decay co-evolve, yet prevailing knowledge tracing models correct bias only after the representation has formed, inherit structure from an incomplete expert matrix, and advance a transition blind to elapsed time. This paper proposes CHM-DTKT, a causal hypergraph selective state space learner twin that treats these as one representation problem. A differentiable hypergraph is built with sampled incidence and an acyclically constrained prerequisite orientation, and its aggregation is backdoor-adjusted over a strictly pre-treatment confounder dictionary, so that deconfounding acts during representation formation rather than at the output layer. A cognitive discretization step generated from elapsed time, item difficulty and cumulative practice induces a per-channel identifiable power-law retention, making forgetting a property of the recursion itself. A predictor–corrector twin decays and repairs a concept-addressable mastery state between observations and supports counterfactual instructional rollout. On five benchmarks under the pyKT protocol the model leads nine baselines by 0.27–0.37 points in the area under the curve, significant after false discovery rate correction at the one percent level on three benchmarks and at the five percent level on the other two, whereas accuracy differences are not significant; the advantage concentrates in sparse and cold-start regimes, with lower calibration error and linear latency growth. Ablation ranks the cognitive step first and representation-level deconfounding second.