During multiple parallel ungrounded Gemini generations on Vertex AI, Treewalker extracts top-5 token-level log-probability distributions for each generated item. When an alternative candidate token surpasses a specified probability threshold (such as 10%), Treewalker splices that token into the preceding prefix to spawn focused completion prompts, discovering latent brand associations. Finally, it aggregates these primary paths and branched completions across runs, computing metrics like appearance rates, mean ranks, and confidence metrics to map the model's parametric memory tree.