@Drunken_Smurf Another possibility is that there's some kind of enforced steering away from reporting the system prompt verbatim (either applied at training or runtime) that results in the variations (tho that doesn't explain Bard. it also thinks it's Bard a lot but then corrects itself...)