Abstract
<title>Abstract</title> <p>K--12 resources in Sichuan and Chongqing span multiple school stages, subjects, and textbook versions across heterogeneous documents. This fragmentation leads to incomplete provenance, uncertain relations, and insufficient subject coverage. Trustworthy recommendation therefore requires both ranking relevance and verifiable evidence for student-facing resources. This study proposes an evidence-graded educational knowledge graph and a trustworthy recommendation framework covering 21 subjects. The framework extracts fine-grained content units with source, page, and textual evidence, distinguishes document-supported structures from candidate relations generated through data augmentation, and links learner state estimation to resource ranking over a candidate graph through risk gating. The resulting graph contains 56,413 nodes and 65,441 edges. Three evidence streams are evaluated separately: knowledge tracing on a public EdNet sample, cross-subject recommendation on learner-disjoint trajectories, and deterministic audits of traceability. For learner state estimation, Attentive Knowledge Tracing (AKT) achieves an AUC of 0.610 and significantly outperforms the classical Bayesian Knowledge Tracing (BKT) baseline after Holm--Bonferroni correction. Risk gating increases the evidence completeness of student-facing recommendations and reveals a quantifiable trade-off between traceable evidence and resource coverage, without statistically significant changes in ranking performance. When no traceable resource is available for a subject, the system returns a resource-gap notice rather than an unsupported recommendation. Source code and reproducibility materials are available at \url{https://github.com/Yizhan-FENG/CYKG-Rec}.</p>