Logprobs evaluate token confidence by taking the mathematical exponent of each candidate token's log-probability to calculate a normalized percentage score. In modules like Treewalker, calculating math.exp(log_probability) * 100 translates raw model logits into an intuitive 0–100% confidence level across generated responses. This metric reveals the model's exact certainty for individual words while systematically exposing alternative branching paths when secondary candidate tokens exceed a designated probability threshold.