The ranker compares candidate edits against competitor items across multiple judge samples to determine the median observed position, awarding a binary reward of one when rank improves and updating the underlying Beta distribution to guide future hypothesis selection.
flowchart LR
A[Candidate Edit Generated] --> B[Ranker Multi-Sample Scoring]
B --> C[Compute Median Rank]
C --> D{Better than Best Rank?}
D -- Yes --> E[Reward = 1, Update Best Rank]
D -- No --> F[Reward = 0]
E --> G[Update Beta: Alpha += 1]
F --> H[Update Beta: Beta += 1]
G --> I[Derive Claim Status]
H --> I