Token-level logprob extraction captures raw logarithmic probabilities directly from Vertex-hosted models like Gemini to calculate exact percentage-based confidence scores for generated brand claims. Implemented within Auxy's Treewalker engine, this mechanism analyzes candidate tokens across independent generation runs to evaluate both primary text confidence and alternative branches that exceed a user-defined probability threshold.

Candidate Logprob → Confidence (%) = exp(log_prob) * 100
             ↳ Threshold Check (≥ 10%) → Spawn Alternative Branch

Because logprob retrieval is incompatible with search grounding tools in Vertex API calls, this technique serves as the core diagnostic for ungrounded, parametric memory audits—identifying weak brand associations and evaluating competitive branch paths before search tools influence the output.