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