Treewalker probes token-level logprobs directly via Vertex AI to map alternative associative paths stored within Gemini's parametric memory. When generating initial brand associations, it inspects candidate tokens that cross a designated probability threshold, constructing a targeted branch-completion prompt for each alternative fragment to generate a full concept without web search. These completions undergo statistical aggregation across repeated runs to quantify appearance rates, average confidence, and hidden associations that standard greedy decoding obscures.