The engine tests snippet variants across multiple rounds using Thompson sampling over a discrete Rank Factor Class taxonomy, applying generative edits and evaluating rank movement through a target re-ranking model.

flowchart LR
    A[Seeder: Hypotheses] --> B[Classifier: Taxonomies]
    B --> C[Acquisition: Thompson Sampling]
    C --> D[Ideator: Variant Crafting]
    D --> E[Ranker: Evaluation]
    E --> F[Posterior Update]
    F --> C