“You can’t prompt-inject a set intersection” — stealing that line. It maps cleanly to an AI ATS checker I’m building: the model can suggest missing keywords, but whether a keyword is actually in the resume should be a plain string match, not the model’s opinion. The referral-export bypass was a great honest find. When perception gets wired in as a soft prior, do you plan a fixed weight against the quiz, or only use it as a tiebreaker?