Judge Model Architecture — DEJAN Codebase Explorer

The judge model is a selected evaluation model configured as the re-ranker in the optimization loop to evaluate competing search snippets or full pages against a target query and output a median ranking over multiple samples alongside per-item rationales.

sequenceDiagram
    autonumber
    actor Optm as Optimizer Loop
    participant JM as Judge Model
    participant Eval as Aggregator

    loop N Samples
        Optm->>JM: Evaluate variant snippets & query
        JM-->>Optm: Return item rankings & rationales
    end

    Optm->>Eval: Aggregate run distributions
    Eval->>Eval: Calculate median rank & variance
    Eval-->>Optm: Final snippet score & next edit priority

Referenced by

Bayesian Optimization Testing Loop
Bayesian Content Optimization Loop
Bayesian Snippet Optimization
Bayesian Acquisition Loop Mechanics
Bayesian Snippet Optimization
Self-Optimizer Knockout Tournament
Bayesian Optimization Loop Details
Re-ranking Model Preferences
Bayesian Content Optimizer
LLM Ranker Evaluation Mechanics