The Bayesian ranker scores each content edit by running the candidate through a multi-sample re-ranker prompt alongside competitor items, extracting both item rationales and a median rank outcome to drive the Bernoulli-trial feedback loop for posterior updates.
flowchart LR
A[Content Variant] --> B[Ranker Model x N Samples]
C[Competitor Pages] --> B
B --> D[Compute Median Rank]
B --> E[Extract Rationales]
D --> F{Improved vs Best?}
F -->|Yes: Reward 1| G[Update Beta: Alpha += 1]
F -->|No: Reward 0| H[Update Beta: Beta += 1]Referenced by