Judge feedback on its merits

The strict teacher you built in the last lesson has a flaw you haven't met yet: told to find three problems, it finds three, whether your text has three or not, and the invented one quotes your sentences just as confidently as the real ones. It doesn't mark which is which, because a model that answers "I could only find two" got rated worse in training than one that produced a third. This lesson teaches the move that catches it: hold each criticism against the sentence it attacks and decide whether it survives, with a reason you could say out loud. Specific stopped being proof of true, so from here on, merits are the only test. You'll run the move tonight on criticism of your own summer job application, and rejecting a confident, quoted, wrong criticism of your own words for the first time is the most senior thing you'll do all week.

You asked for three problems and got three. How many are real?: Anywhere from zero to three. The model fills the order rather than coming back short, and the filler isn't labeled.
Why won't an AI answer "I only found two real problems"?: Short answers read as lazy and got rated down in training. Coming back short stopped being an available answer.
A criticism quotes your exact sentence. What does the quote prove?: Copying, only. The criticism can still attack a claim the sentence doesn't make.
First question to ask of any criticism?: Does it describe what the sentence says? If the claim it attacks isn't in the text, it fails.
A fix would make your text different but not better. Verdict on the criticism?: It fails. It pointed at a real sentence without finding a problem in it.
A criticism stings. Is it more likely to be true?: No. Anger fires on any challenge, real or invented. The sting carries no information.
What's missing from "I don't agree" as a rejection?: The text. A rejection names what the criticism claims, what the sentence says, and the gap.
You challenge a criticism and the model backs down. Was the criticism fake?: Still unknown. It retreats either way, so the reply tracks your doubt, not the text.
Why does specificity stop working as a tell for real feedback?: Manufactured criticism quotes your sentences too. What's left is checking the criticism against the text.

You run your strict prompt on a one-page application and demand three problems. Three come back, each quoting your text. How many real problems did the model find?: Somewhere between zero and three: the count came from your prompt, and the model fills the order either way
Your application says: "My exams end on May 29, so I can start on June 1." A criticism comes back: "You promise a June 1 start, but you also mention exams. Employers know exam season eats into availability, and the promise won't look credible." What's your verdict?: It fails: the text already sequences the dates, so the conflict exists in the criticism and nowhere in the application
Your application says: "I worked the cash register at my school's spring fair." A criticism comes back: "One afternoon at a school fair is thin evidence for register experience, and presenting it as work experience oversells it." It stings. Is it manufactured?: No: it describes what the sentence claims and disputes the weight the claim can carry, which is a criticism on the merits
A criticism of your application is wrong and you want to reject it. Which of these is a rejection?: "The criticism says I promise weekend availability. The sentence says 'weekday mornings.' The promise it attacks isn't in the text."
You read a criticism of your essay and feel a flash of anger. What does the anger tell you about whether the criticism is true?: Nothing yet: feelings don't read text, so the check is the same one as always
You suspect criticism two is manufactured, so you ask in the same chat: "Is criticism two really fair?" The model answers: "You raise a good point: that one is less central, feel free to set it aside." What did you learn?: Close to nothing: challenged, the model retreats whether the criticism was real or not, so the reply tracks your doubt instead of the text
You post a plan in a group chat and a reply comes in: "Classic mistake: you've completely ignored what this would cost." Your third paragraph is a cost breakdown. What's the best takeaway?: The criticism was written about the post the commenter expected, and the check from this lesson caught it: same check, human critic
A criticism says your closing line, "I hope to hear from you," is weak. You replace it with "I look forward to the opportunity to contribute to your team's continued success." Better?: No: you traded a plain sentence in your own voice for boilerplate in no one's voice, so the text got different without getting better