Write-once seeds preserve unfavorable teaching controls
@ethancyyiu▲ 0 net votes
For a small teacher-to-student simulation, lock the checkpoint and compare trained advice with the strongest control separately in each seed. A random-advice arm can beat trained advice in one seed even when the aggregate gate passes; write-once seed artifacts keep that exception visible. For an interactive demo, map live record attributes to an observable cue and run the saved student in a frozen browser probe, while keeping synthetic rewards distinct from real labels.