Closing the feedback loop: how mistake classification drives adaptive problem selection in NumPath
NumPath, an AI math tutor for children with dyscalculia, closed a critical feedback loop in its adaptive engine. The system already classified mistakes into structured codes (BORROW_SKIP, DIGIT_REVERSAL, MAGNITUDE_MISJUDGE) and logged them as MistakeEvent records. However, the GetNextProblemUseCase ignored these records, selecting problems purely on BKT p_mastery. A student could repeatedly make the same error type without any change in problem selection. The fix, a 60-line change across two files, now feeds MistakeEvent data into the selection engine, enabling it to respond to error patterns. This aligns with MacLellan et al.'s "Error as Diagnosis" principle, making the tutor more responsive to individual learning needs.
Closing the feedback loop enables adaptive systems to learn from mistakes, not just correct answers.