In human-in-the-loop page mode, the system seeds and classifies hypotheses before pausing, allowing the user to drive optimization by submitting manual patch edits and factors while the ranker tests each variant and updates the Bayesian posterior.