ECLI-Agent: A Retrieval-Augmented LLM Framework with Explainable and Counterfactual Reasoning for Personalized Intervention in Sustainable E-Learning

Counterfactual actionability: the proposed Counterfactual Actionability Test (CAT) moves beyond feature importance by determining whether a feasible change in an actionable learner feature can produce a meaningful reduction in predicted risk.
2) Verified retrieval-grounded intervention generation: intervention language is generated from retrieved evidence, while counterfactual targets, risk values, evidence identifiers, and final verification remain outside the LLM and are handled deterministically.
3) Evaluation beyond predictive accuracy: the framework is assessed through full-cohort counterfactual actionability, subgroup analysis, Direct-LLM versus RAG+LLM comparison, and a frozen unseen evaluation with fallback and abstention.