Logprob confidence scores are calculated by converting the model's raw token-level log probability values into percentage confidence metrics, recording both the chosen sequence and the probabilities of top alternative candidate tokens.