The Bayesian optimization loop begins with a seeder generating testable hypothesis claims with Beta priors, which are categorized into hierarchical Rank Factor classes. Across iterative rounds, a two-stage acquisition policy balances class freshness and Thompson sampling to pick claims, directs an ideator model to craft isolated snippet or line-numbered page edits, and measures median rank shifts across multiple judge samples to update posterior beliefs until reaching rank one or exhausting the round budget.