Treewalker audits an AI's parametric memory by analyzing token-level logprobs across repeated ungrounded generations and tree-walking alternative paths above a probability threshold to surface hidden brand associations.
sequenceDiagram
autonumber
participant Engine as Treewalker
participant API as Vertex AI (Gemini)
Engine->>API: Send base prompt (ungrounded)
API-->>Engine: Return primary tokens & logprobs
rect rgb(30, 40, 60)
note over Engine: Filter alternate tokens (P > threshold)
loop For each high-prob candidate
Engine->>API: Request branch completion
API-->>Engine: Return branch tokens & probabilities
end
end
Engine->>Engine: Aggregate brand frequencies & logprob weights
Engine-->>Engine: Output parametric visibility matrix