Each optimization claim starts with a Beta prior distribution derived from an honest seeder prior, which is iteratively updated by incrementing posterior parameters based on the binary Bernoulli trial outcome of whether the edited variant improved search rank.

stateDiagram-v2
    [*] --> Initialized: Seeder assigns Prior P(H)
    Initialized --> Selection: Beta(alpha, beta) defined
    
    state "Testing Loop" as Loop {
        Selection --> Evaluation: Thompson Sampling
        Evaluation --> Reward: LLM Ranker (Before vs After)
        Reward --> PosteriorUpdate: Binary Reward (r=1 or r=0)
        PosteriorUpdate --> Selection: Update alpha += r, beta += (1-r)
    }

    PosteriorUpdate --> Classification: Convergence / Threshold Check

    state Classification {
        [*] --> Untested
        Untested --> Testing
        Testing --> Likely: High Win Rate
        Testing --> Weak: Low Win Rate
        Likely --> Proven: Strong Significance
        Weak --> Disproven: Strong Rejection
    }

    Classification --> [*]