Ways to watch the same detectors work: the researchers' run codes, Bayesian Knowledge Tracing, the original and discovered clusters, and timed feedback suggestions.
The researchers' rule-defined groups and the LLM-discovered clusters, side by side: what each group is, why each scheme exists, where they overlap and differ, and how the students in each overlap did.
2 · One student at a timeSet up runs by hand or replay examples, and watch a single student move between groups on both sides, with an explanation of each estimate and cautions where a reading misleads.
Replay real sessions with both clusterings, Bayesian Knowledge Tracing and timed feedback suggestions; pick students by original group, discovered cluster or BKT trajectory.
Bayesian Knowledge TracingHow BKT estimates each skill run by run, with the parameters exposed.
The real modelUse the actual Wildfire model while the detectors listen to its log events and suggest the next step.
The full analysis behind these demos. Open any document here, or download it.
How the pieces relate. The codes and rule case are exact, computed the way the researchers compute them. BKT estimates how well a student has each skill right now. The original grouping (no engagement / exploring / control of variables) comes from the researchers' streak coding; the discovered clusters come from the LLM analysis and are estimated live by a small model (86% accurate on whole sessions). The and the documents above give the details.