Treewalker extracts token-level log probabilities and alternative candidate tokens directly via Vertex AI to expose latent associative paths in Gemini's parametric memory. During initial generation passes, it captures the model's top alternative predictions, combining the prior token prefix with each candidate that crosses a specified probability threshold to construct a branch fragment.

Concurrent worker threads then dispatch targeted completion prompts for every identified fragment, instructing the model to complete the alternative phrase into a full concept. The engine compiles these branch trajectories across multiple runs, statistically aggregating appearance frequency, mean rank, and token confidence scores to illuminate associations that standard decoding passes bypass.