The system balances exploration and exploitation by prioritizing unproven hypotheses that hold high upside variance while continuously refining winning edits, combining probabilistic belief updates with class-level freshness weighting to prevent premature convergence on a single rank factor.
flowchart TD
A[Start Round] --> B{Stage 1: Class Freshness}
B -->|Sample by Weight| C[Rank Factor Class]
C --> D{Stage 2: Thompson Sampling}
D -->|Draw θ ~ Beta per Claim| E[Highest θ Hypothesis]
E --> F[Ideator Crafts Edit]
F --> G[Ranker Scores Variant]
G --> H{Rank Improved?}
H -->|Yes| I["Update α (Success)"]
H -->|No| J["Update β (Failure)"]
I --> K[Update Posterior & State]
J --> K
K --> A