Generative AI could turn learning analytics into personalized support for programming students, but such systems still need rules for when to act. Feedback tools must decide whether to intervene, what data to draw on, and how much help to provide.
The paper describes a "risk-adaptive" framework that sets these boundaries. Because it is risk-adaptive, the system changes its behavior based on the perceived risk of a learning situation. Because it is evidence-constrained, it selects a defined set of evidence rather than using everything collected.
The research addresses a practical gap: without clear constraints, AI feedback may be too frequent, under-supported, or overwhelming. This framework offers a structure for balancing support with student autonomy.