Treewalker queries Gemini over Vertex AI without search grounding, capturing token-level probabilities across the top five candidates per position by converting logarithmic values (\(e^{\text{log\_prob}} \times 100\)). Whenever an unselected candidate token exceeds a designated probability threshold, the engine tree-walks that alternative prefix through secondary completion prompts to systematically map hidden associative paths across multiple independent runs.