Treewalker investigates a model's internal representations by querying ungrounded models across multiple runs to map brand associations without the influence of external retrieval. For every response, it extracts token-level logprob confidence scores to calculate average and minimum probability metrics, and then executes a tree-walking high-probability alternatives algorithm on tokens exceeding an established confidence threshold. By recursively prompting the model to complete alternative token fragments, the engine discovers latent brand associations and calculates aggregate appearance rates across iterations.