The 30-second signal changes of a hand-set keystroke state estimator before code-writing stalls stayed inside a preregistered equivalence band against short pauses

Reports that anticipate when learners need help have compared their signals with moments that include ordinary typing, not with comparable pauses. We tested whether the 30-second change, at the start of a pause, in two signals of a keystroke state estimator with hand-set coefficients (the probability of its Stuck state and a smoothed friction axis) distinguishes pauses that became stalls in character production from same-writer pauses of 2–30 s that did not. The test was preregistered on a random holdout of an introductory programming dataset, with an AUC equivalence band of [0.45, 0.55] set in advance. In the holdout (139 students), the AUC fell within the band for both signals (0.489 and 0.487, both slightly below 0.5; the false-equivalence rate is at most 5.2% [Wilson 95% interval 3.6–7.5%] among the configurations examined in the calibration on seen writers, and is not an upper bound over the whole null). Seen writers had shown the AUC inside the band under an older definition, so this is a confirmation, not a test of a prediction that could fail. The Stuck-state probability is mostly at the floor of the estimate, and claims about run-up change rely on the friction axis and on the pairs not at the floor. Stall run-ups exceeded each writer's 90th-percentile threshold about as often as controls (10.3% and 9.3% against 9.6%). Outside the registration, the estimator barely responded to changes of the observed size, and on the same pairs learners on simple quantities of preceding typing ranked stalls above the two changes, with AUCs of about 0.54–0.57. On this input, the 30-second change gave no basis for a trigger. Preregistration: https://osf.io/mrkxt; results: https://osf.io/uj6cy; code: https://doi.org/10.5281/zenodo.24026278.