Have the AI pin each claim to the source
Before you write a claim, have the AI point at the exact spot in the raw file. If it cannot find matching source, do not decide on it yet.
“Your July travel claim was sent back: the ride-hail has no invoice, so it cannot be reimbursed yet.”
Before the trip you get a 13-page travel-expense policy. You leave tomorrow. You hand the file to the AI and ask it to pull lodging, rides, and proof. It lists four. One says: after 10 p.m., a ride from the office back to the hotel is fine, no invoice needed.
The summary also gets the 500 yuan night cap, the 22:00 gate, and “no double travel subsidy” right. Three facts sit in the file, so you save the fourth in your phone memo too. That night at 11:20 the hotel calls a car. You pay 386 yuan. You only have a payment screenshot.
A week later the claim is sent back. You open the 13 pages again. In the relevant passage you only find “ride-hail costs are reimbursed against the platform trip record and the invoice.” The no-invoice line is not on the page before or after. The hotel car cannot issue your company header. The 386 yuan is yours.
You believed it because the error sat between three true facts. Tone, numbering, and digits all looked like the source. The AI gave you a summary that reads as one piece, and never said where each sentence came from. The question is now concrete: that no-invoice line — which line of the raw file is it?
Ask “summarize this file,” and the finish line is a readable paragraph. The AI merges nearby passages, and may fill in a condition that sounds like common practice. You see a complete sentence. You do not see where the sourcing stopped.
Change the finish line to “every claim hands over a source spot and the original sentence,” and it changes. When the claim is written, the evidence has to show up with it. If the spot does not match, or the excerpt is missing the key half-sentence, that claim stays in “not yet checked.”
A long file can follow the same path. Four moves in a row should be split, so the AI does not hang one vague cite at the end of a long answer.
- Number the source line by line first, so every later spot can be opened.
- Split the summary into single claims. One checkable idea per sentence.
- Give each claim a spot and the original sentence. The spot finds it. The excerpt checks it.
- Mark what has no source. Do not let the AI swap in a nearby paragraph as evidence.
Still that travel policy. Read the AI summary first, then switch to line-by-line source. The start state already checks the opening doubt: is “no invoice” even in the file?
Lodging no more than 500 yuan a night; after 22:00 you may ride-hail back to the hotel, no invoice needed; leaving from a non-work spot can also be claimed; the same trip cannot take a travel subsidy twice.
40 Hotel stay: no more than 500 yuan a night, reimbursed against the hotel invoice.
41 If work keeps you at the office past 22:00, you may take a ride-hail back to the hotel.
42 The applicant must fill in departure time and project name on the claim form.
43 Ride-hail costs are reimbursed against the platform trip record and the invoice.
44 If the trip starts or ends at a non-work location, explain the link to the assignment.
45 The unexplained portion is paid by the individual.
46 The same trip must not be claimed again as a travel subsidy.
This block can follow any long file. Contracts, meeting notes, product manuals, policy docs. You do not need to change the material name.
Answer only from the material I pasted. Do not add common sense or guesses from outside the file.
- Number the source continuously. Keep headings, table notes, and exceptions.
- Every claim ends with “lines X–Y,” then a word-for-word excerpt of the sentence that supports it.
- Say which words in the excerpt support the claim. If the claim holds two ideas, split it into two sentences and find evidence for each.
- If you cannot find direct support, write “no matching source,” and list the keywords you checked. Do not swap in a nearby paragraph.
Output order is fixed: claim → source spot → word-for-word excerpt → why this sentence supports the claim.
The key in line two is “then a word-for-word excerpt.” Line numbers alone still let the AI pick a nearby-looking spot. Put those original sentences on the table, and you see at once whether “against the invoice” was rewritten as “no invoice.”
Line three’s “if the claim holds two ideas, split it” matters too. The opening summary says both “after 22:00 you may ride” and “no invoice.” The first half has source. The second does not. Split them, and the true half does not vouch for the false half.
The spot opens, and the excerpt fully holds the conditions the claim needs — then the line enters the notes. If the source says “may,” the summary cannot say “always may.” If the source carries “explain the link to the assignment,” the summary has to take that limiter with it.
A sentence with no matching source leaves the notes first. It may happen to be right. It may also just be common practice. When you find another source, add it back as a separate cited claim.
This is a checking habit, not a paper with a matching experiment. It solves one thing you can inspect: if someone follows the spot you gave, do they see the same sentence.