We present a generative‑AI feedback framework for introductory programming that integrates risk‑adaptive intervention, evidence‑constrained generation, and capacity‑limited policies. Using 2,993 failed‑submission states from 215 students, we built student‑disjoint predictive models to forecast persistent failures and related outcomes.
Four matched feedback conditions were generated for 136 cases, and calibrated risk scores guided the timing of interventions. The selected logistic‑regression model achieved $\text{PR‑AUC}=0.550$ and $\text{ROC‑AUC}=0.681$ on the test set. Incorporating richer student histories further improved prediction of unmodified resubmissions.
After standardized repair and evidence gating, 519 of 544 newly generated messages contained all required components. A fixed‑threshold sequential policy selected 17.8% of eligible test states and captured 25.2% of observed persistent failures.
These findings endorse an evidence‑gated progressive assistance strategy: calibrated risk determines when to intervene, recorded evidence constrains feedback content, and assistance progresses from self‑checks to localized hints as needed. The framework links prediction, decision‑making, and grounded generation while keeping their evaluation outcomes distinct.
Review