In subsequent rounds, the ranker's per-item rationales are fed directly into the subsequent ideator turn as contextual feedback, enabling the model to refine its edits against specific competitor strengths and weaknesses while updating the Bayesian posterior.

flowchart TD
    subgraph Ideation["Optimization Loop"]
        Claim["Hypothesis Claim"] --> Ideator["Ideator"]
        Draft["Content Draft"] --> Ranker["Ranker Evaluation"]
        Ranker --> Rationales["Per-Item Rationales"]
        Ranker --> Posterior["Bayesian Posterior Update"]
        Rationales --> Ideator
        Ideator --> Draft
    end