Logprob thresholds act as a statistical filter applied to the model's top-5 candidate logprobs at every generated token position, establishing the minimum probability required for an alternative token to be pursued. When an alternative meets or exceeds this threshold percentage, Treewalker uses that token prefix to trigger separate recursive branch-completion calls, systematically surfacing latent, competing associations and variations that the model weighed internally during parametric recall but ultimately left out of its final greedy output.