Ungrounded hallucinations and legacy associations are flagged when concepts surface with high confidence or appearance rates in unprompted parametric recall but fail to appear in live search results or get contradicted during grounded verification.

flowchart TD
    subgraph ParametricAudit["Ungrounded Model Probing"]
        P1["Brand Association Prompt"] --> P2["Treewalker Logprob Analysis"]
        P2 --> P3["High Confidence Output<br/>(Appearance Rate >= 80%)"]
    end

    subgraph GroundedAudit["Grounded Live Search"]
        G1["Live Web Search Query"] --> G2["Fact & Grounding Inspector"]
        G2 --> G3["Zero Corroborating Sources<br/>(Empty Search Chunk)"]
    end

    P3 --> FGC["Fact & Gap Checker"]
    G3 --> FGC

    FGC --> V1["Verdict: Ungrounded Hallucination"]
    FGC --> V2["Verdict: Stale Legacy Association"]

    V1 --> ACT["Trigger Bayesian Content Optimizer"]
    V2 --> ACT