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 --> ACTReferenced by