This formula converts the model's negative log-likelihood into a 0–100% probability score for each emitted token, allowing Treewalker to measure confidence and compare alternative candidate tokens against a branching threshold.