Replay a real student's runs one at a time and watch the detectors respond: the researchers' codes, BKT mastery, a live cluster estimate, and the feedback a coach would suggest at each moment.
drought-runs.json (made by scripts/export_replay.py; in data/replay/ on the laptop and the mini). It is read in this browser only and never uploaded.N nothing differs · O only drought differs · I one other setting · M several. Upper case = one spark per zone. = identical rerun · * zones same, wind/sparks changed · + same comparison, tweaked.
Shown when the replay reaches the last run, so the live estimates can be judged first.
The original side is computed exactly from the runs. The discovered side is a logistic model on rule-computed features that predicts the Sonnet-labelled cluster: 86% accurate on whole sessions (cross-validated), 70% after 3 runs. BKT, above, tracks skill; both sides here name a pattern of behaviour.
Timing rules: a new rule case shows its round-1 line (original side, dark edge); the same case on a second attempt shows the discovered cluster's coaching focus (round 2, ember edge); BKT adds an early-alert nudge and a mastery affirmation (green edge). Round-1 lines are study-written Hazbot-style templates.
| # | Zone 1 | Zone 2 | Sparks | Wind | Code | Fair test | P(L) after |
|---|