A friend who does industry research showed me his draft. Forty-seven citations, each with a source attached.

His method, the night before delivery: open all forty-seven links and confirm from memory that each one said roughly what he had claimed.

The problem with that process isn't speed. It is that it doesn't check anything.

Confirming from memory is not confirming

People are confident about their own recall in a way that isn't earned. The feeling that "I have definitely seen this line" usually means it fits what you expected to find, not that you read it.

Three wrong out of forty-seven, a reviewer picks those three, and the whole document is compromised. He had no way to guarantee he would get all forty-seven right.

We tried letting the model judge

Version one: the model reads the source, reads the quote, decides whether the source supports it.

It didn't work. Not because it was wrong, because it was unstable. Same passage, two runs, two answers. A longer source and it starts skipping sections. A scanned PDF and it invents a paragraph.

One false "supported" and the feature is worth nothing, because you can no longer skip checking.

Splitting the job in two

  • Locate. Where in the source does this quote come from? A rule does this: string matching, context windows, paragraph offsets. If it can't find it, it says it can't.
  • Judge. Does that passage support the claim? The model does this, but only after a location exists, and it has to show you the passage it is reading.

So each note reads like this: paragraph 3, sentence 2, "significantly improved". Source (page 12, paragraph 4) says "improved somewhat". Your wording exceeds the source. Click, and you are on that page.

Nine of the forty-seven had problems

Six were overstated wording. Two had no matching sentence in the source. One cited the wrong section.

None of them were "the AI thinks you are wrong". It only reports things it can point at.

He said something afterwards that I have kept: I don't need it to tell me whether I am right. I need it to tell me which sentence to go back and look at.

Related: Why the AI can't look things up for you · GB/T 7714 and APA: where auto-generation should stop