Logprob distributions represent the logarithmic probability values assigned to candidate output tokens at each generation step, allowing diagnostic probes to quantify model certainty and uncover hidden semantic associations through multi-run sampling.