During multi-sample re-ranking, the judge model evaluates the full set of candidate snippets multiple times per round to derive a stable median rank and extract rationale from the canonical sample.

sequenceDiagram
    participant O as Optimizer Loop
    participant R as Ranker Model
    participant S as Statistical Median
    O->>R: Sample 1: Evaluate candidate + competitors
    O->>R: Sample 2: Evaluate candidate + competitors
    O->>R: Sample N: Evaluate candidate + competitors
    R-->>S: Return N ranking orders & rationales
    S-->>O: Compute median rank & canonical rationale