The LLM ranker acts as an objective referee by evaluating candidate snippets or full pages across N samples using a chosen judge model, recording both the median rank and structured rationales for why items placed higher or lower than competitors.

sequenceDiagram
    autonumber
    participant Cycle as Optimization Loop
    participant Ranker as LLM Ranker
    participant Judge as Judge Model (LLM)
    participant Agg as Rank Aggregator

    Cycle->>Ranker: Send variant + competitor set
    par N Parallel Samples
        Ranker->>Judge: Prompt pairwise/listwise evaluation
        Judge-->>Ranker: Return rank order + rationales
    end
    Ranker->>Agg: Aggregate N evaluation samples
    Agg-->>Ranker: Compute median rank + structured rationales
    Ranker-->>Cycle: Return ranking trace + posterior update