"Interestingly, the base pre-trained model is highly calibrated (its predicted confidence in an answer generally matches the probability of being correct)."

Is that the same confidence measure you can tease out by prompting "to each of your statements output your estimated confidence in it's truthfulness" ?