Treewalker queries Gemini over Vertex AI with logprobs enabled, evaluating the top candidate token distributions to quantify the model's baseline confidence in each recalled association. Whenever alternative tokens exceed a designated probability threshold, the engine captures the preceding prefix, splices in the competing candidate, and triggers branch completions to map out adjacent latent concepts. By executing multiple parallel sampling passes alongside search-grounded baseline runs, the system calculates appearance rates, rank variance, and average token probabilities. This comparative pipeline exposes stale training associations and latent brand perceptions that standard single-pass generation fails to surface.