Page:Line or It Didn't Happen: Making AI Cite the Record

The single most dangerous thing an AI will do in legal work isn’t get a fact wrong. It’s get a fact plausibly wrong — assert something the record never said, in clean, confident, lawyerly prose, with no flag that it was invented. That’s the failure that ends up in a brief.

The fix isn’t a better model. It’s a discipline you can impose on any model: page:line or it didn’t happen.

The rule

Every factual claim the AI makes about the record must point to where it lives — 47:12, 112:3–113:8. If it can’t cite, it doesn’t claim. And when the answer isn’t in the record at all, the correct output is “not in the record,” not a confident guess.

That sounds obvious. It is. It’s also the thing almost no off-the-shelf legal AI actually enforces, because it’s easier to demo a slick summary than a disciplined one.

How to make a model do it

You don’t need a special tool — you need the prompt to require it:

• Give it the record as the only source of truth. “Answer only from the transcript provided.”
• Demand citations inline. “After every factual statement, cite the page:line. No citation, no statement.”
• License the non-answer. “If the transcript does not address the question, respond exactly: not in the record.”
• Ask it to quote, then characterize. Pulling the verbatim line before summarizing keeps it honest.

This is the backbone of the CITE method, and it’s the entire reason CasePrompts exists: copy-ready litigation prompts that bake page:line discipline in, so you’re prompting like you’re assigning a task to a junior associate — not Googling. Because the discipline lives in the prompt, it’s portable: the same one works across whatever model you’re using this month.

A model that cites the record is a research assistant. A model that doesn’t is a liability with good grammar. The line between them is a prompt.

Want the exact prompt I use for transcript Q&A? It’s in the CasePrompts library — and I’ll walk through it line by line if that’s useful.