An Answer Without a Citation Is a Guess

Point a document-reading system at a stack of contracts and it will answer your questions. What is the concentration limit for this fund? What did the third amendment change? The answers come back fluent and fast, and most of them are right. That is exactly the problem.

In finance and legal work, “most of them are right” is a liability, not a feature. The value of these answers is not the ninety-odd percent that are correct. It is the handful that are wrong in a way nobody catches: the misread limit that clears a trade it should have blocked, the amendment the system quietly ignored. You cannot manage a risk you cannot see, and an answer with no source is an answer you cannot check. It looks identical whether it is right or wrong.

The people who work in these domains already know this. It is why a credit officer wants the page, an auditor wants the tie-out, and an attorney will not put a fact in a filing they cannot trace to a document. The standard was never “the system seems confident.” It was always “show me where this came from.”

So build the system to that standard.

Do not let a fact exist unless it is tied to its source. The rule is simple and it is architectural: a claim enters the record only as a statement anchored to a specific document, a specific version, a specific page, and the exact span of text it came from. Extraction does not read a number and store it. It reads a number and stores it with a pointer back to where it read it. There are no floating assertions, because the schema does not allow one.

Make the assistant cite, or decline. An assistant that has read the documents should answer with the citation attached, so checking it means opening a page, not re-deriving the analysis. And when it cannot find support, the correct behavior is to say so — “I do not have a source for that” — not to produce a fluent sentence that fills the gap. A system that guesses gracefully is worse than one that admits the gap, because the graceful guess is the one that ships into a decision.

This is a different posture toward the model than demo culture rewards. The demo optimizes for the impressive answer. Production in an operation that gets audited optimizes for the checkable one. Citations are what convert “trust the model” into “check the source”. A ten-second check is a thing a busy operation will actually do, where a leap of faith is a thing it eventually regrets.

The citation is also the audit trail. When every fact points at its origin, the record of the analysis includes the record of where each piece came from. A question about a number six months later has a clean answer: here is the document, the page, the span, and the version it was current as of. Nobody reconstructs the reasoning from memory, because the reasoning never left the evidence.

None of this is free. Anchoring every extracted fact to a page is more work than piping documents through a model and printing the output. Page-level segmentation, version tracking, and span-level references are real engineering, and they slow the impressive first demo down. That is the trade, and in finance and legal it is not close: the easy version is the untrustworthy one.

Two GrayBeck systems are built on this rule. In the Avail suite, document intelligence reads credit agreements and borrowing-base packages and extracts terms anchored to the page they appear on, and the assistant that answers questions about them cites the page it relied on. In MatterOS, every cell of a matter’s analysis grid cites its evidence down to the document, page, and span, and nothing exists in the record as an assertion an attorney cannot check against its source. Different domains, same discipline.

The test is easy to apply to any system a vendor is selling you. Ask it a question, then ask it to show you where the answer came from. If it can open the page, you have a tool you can check. If it cannot, you have a confident narrator. In work that gets audited, a confident narrator is the most expensive thing you can buy.