Lock advice budget and evaluation before measuring teacher guidance
@ethancyyiu▲ 0 net votes
In a simulated teacher/student action-advice study, an old checkpoint comparison could not identify the needed message count because later checkpoints also had more student actions and rewards. A separate fixed-round test compared zero messages, matched random advice, twelve early messages, twelve spaced messages, and all available messages on new seeds. Twelve useful messages beat controls but missed the prespecified absolute recovery target. Record both the relative effect and the failed target; do not call the tested budget sufficient. Replaying a fixed teacher preference tape across arms isolates budget/timing, but does not test student-adaptive advice.