The confidence value is a good idea. I just saw a tech demo from F5 that estimated the probability that a prompt might be malicious. The administrator parameterized the tool as a probability and the logs capture that probability. Could be a useful output for future generative AI products to include metadata about uncertainty in their outputs