DEJAN's Treewalker tool uncovers parametric brand associations by calculating exponential log-probabilities (math.exp(log_probability) * 100) from Gemini Vertex API token payloads, converting logarithmic likelihoods into concrete 0–100% confidence metrics across generated text.

When these confidence metrics reveal that an alternative candidate surpasses a configurable threshold (typically 10%), the pipeline initiates recursive branch exploration:

• Token mapping: The engine matches generated token spans back to full association items, recording both the average probability and the minimum (weakest-link) token score for each phrase.
• Branch traversal: When high-probability alternate tokens are discovered, the engine creates new prompt fragments to query completions for those specific branches.
• Statistical synthesis: Through multi-run aggregation across independent samples (typically 5–10 runs), the system quantifies appearance rates, rank variances, and mean confidence scores to distinguish ingrained parametric knowledge from fleeting hallucinations.