Token-level logprobs measure the exact conditional probability assigned by the model's vocabulary distribution to each generated token, revealing how deterministically an association is recalled versus competing semantic alternatives.