Evidence, methods, and open questions.

Offline replay on public student logs.

ARIA-Pace was fit on public, de-identified 2009–10 ASSISTments logs and evaluated once on a locked held-out split of 597 test students. When interruptions are weighted three times as costly as a delayed hint (the Quiet cost setting), under the 3-correct-in-a-row mastery definition, it intervened 52% less often than helping after every wrong step and missed 3.7% of students who went on to get stuck (95% CI 2.2–5.3%).

When a delayed hint is weighted three times as costly (the Supportive setting), the saving shrinks to 3% and the difference to always helping includes zero. Quiet, Balanced and Supportive are cost settings we chose, not types of learner. This is a retrospective policy replay, not a classroom study or evidence that an intervention improves learning.

Sources: the specification, code and result files are in the aria.kls repository (branch aria-pace: docs/aria_pace/RESULTS.md and aria_pace/results/test_results.json). The protocol, validation numbers and limits are on the ARIA-Pace page.

Development benchmark

Language classification under stress.

These numbers describe the language classifier on synthetic examples. The demo’s decision to step in after a wrong step comes from the math checker and the timing model, not these labels; the labels can only add optional flags that need an explicit feeling word.

Specific ways to collaborate.

• Review the proposed school consent protocol and research ethics.
• Independently label appropriately permissioned student reasoning transcripts.
• Review intervention costs, replay assumptions, and statistical methods.
• Help design an approved feasibility study at a school.

Please send a description of a potential dataset first, without attaching student records or identifiable transcripts.