Cite or Refuse: How Rexfin's AI Answers Without Ever Guessing

Ask a general-purpose AI assistant a numeric question about a company and it will answer. That’s the design. It’s trained to be helpful, and a fluent non-answer feels like a failure to the model, so it fills the gap with something plausible. In finance, a plausible number that isn’t real is worse than no answer at all, because it looks exactly like a correct one until someone checks it against the filing and finds it doesn’t match.

Rexfin’s answer engine is built around a rule that sounds simple and is deliberately hard to route around: every numeric claim in an answer has to resolve to a citation, or the engine refuses to answer that part of the question. Not “flags it as uncertain.” Refuses.

Two jobs, split on purpose

The engine separates finding evidence from writing the answer. A first pass (an investigator) runs a multi-step search over the graph of extracted facts and their relationships, looking for the specific data points the question actually needs. Only once it has gathered that evidence does a second pass (the writer) draft the answer, and it’s constrained to write only from the evidence block it was handed. It cannot reach past that block into whatever it happens to “remember” about the company from its training data. If the evidence doesn’t contain the number, the writer doesn’t have access to a version of the number that came from anywhere else.

That constraint alone isn’t the whole safeguard, because a model can still misstate what its own evidence says. So there’s a separate, deterministic check afterward: a citation-truthfulness gate that looks at every number in the drafted answer and verifies it’s actually backed by the cited evidence, not just plausible next to it. A number without a real citation behind it doesn’t get softened with a caveat: the answer is refused, or the unsupported part of it is dropped and the response says so plainly. A verification pass runs on top of that and returns a confidence label (high, medium, or low) so a reader isn’t just told an answer was checked, but how strongly.

Refusing usefully, not just refusing

A flat “I can’t answer that” is nearly as unhelpful as a wrong number. So when Rexfin can’t answer a question as asked (because it’s ambiguous, or scoped to more than one filing, or asks for something the underlying documents don’t contain) it says what it can answer instead, rather than leaving a dead end. If a question is ambiguous about which company or filing it means, the engine asks for the missing scope instead of guessing which one the user probably meant.

Scope discipline runs deeper than that. A number that belongs to one reporting segment doesn’t get quietly cited as if it were the whole company’s figure, and a question scoped to one fiscal year doesn’t get answered with a number pulled from a different year’s filing just because it happens to be the newest one on file. Those sound like small distinctions, but they’re exactly the kind of mix-up that turns a confident-sounding answer into a wrong one that’s hard to catch on a quick read.

Where the evidence itself comes from

None of this refusal discipline would mean much if the evidence pool it draws from was shaky. The citations the writer is constrained to come from figures that already passed reconciliation to the filing’s printed totals and sit at the appropriate rung of the trust chain: a filed, verified figure is treated differently from an internal estimate, and the answer engine respects that distinction rather than flattening it. That’s also why the citations behind an answer link straight into the provenance drawer, so a reader can go from the sentence to the source document in one click instead of taking the answer on faith.

Who this is built for

This matters most for anyone who has to stand behind an AI-sourced number in front of someone else: a CFO briefing a board, an analyst answering a lender’s question, a controller fielding an auditor’s follow-up. If the cost of a wrong answer is your own credibility, “confident but ungrounded” is the failure mode you can’t afford, and cite-or-refuse is the design that removes it. Explore the rest of the Rexfin product tour for how the evidence layer underneath this actually gets built.