How Verification Works: The Check Behind Every Exported Number

A number in a board pack is only as good as your ability to defend it if someone pushes back. Most AI tools skip that part: they narrate a figure fluently without checking whether it’s consistent with anything else in the model. Rexfin’s verifier closes that gap: before a figure is exported, it’s checked against the accounting identities that have to hold for it to be real, using exact arithmetic rather than an approximation.

Every number has to be anchored first

Before any identity gets checked, every numeric claim has to be traced to a specific source. Rexfin calls this anchoring: a number gets matched to the underlying fact it came from (the fact behind the citation in the provenance record) by value, by what it represents, and by the period and entity it applies to. Two line items that share a value don’t cross-match just because the numbers line up. If a number can’t be matched to anything, it’s marked unverifiable, and on the answer side that’s a hard stop: the sentence gets refused rather than shipped, the same posture as how ingestion works, one layer deeper into the arithmetic. A figure also needs corroboration from more than one independent read of the source page before it’s cited outright; a single-read match is shown with lower confidence and can’t satisfy a hard check alone.

Two kinds of numbers, two kinds of failure

Once numbers are anchored, verification checks whether the accounting identities that should hold actually do: assets equal liabilities plus equity, and operating, investing, and financing cash flow plus the currency effect nets out to the actual change in cash. These aren’t new checks: they’re identities every accountant already knows have to be true. What’s new is that Rexfin actually runs them, on every relevant period, instead of leaving them as a claim nobody enforces.

The result depends on where the numbers came from. If every operand traces back to an actual figure from the filing, a failure is treated as real and blocks the export: you cannot ship a board pack whose balance sheet doesn’t balance. If any operand is a modeled or forecast figure instead, a failure is only ever a warning. Projections are allowed to be wrong in the way projections always are; audited actuals are not.

Exact math, not float approximation

Financial statements are printed rounded, to the nearest thousand, sometimes million. A verifier checking for exact equality would fail every correctly prepared statement on trivial rounding, and one using ordinary floating-point math can introduce its own small errors on long sums of large numbers. Rexfin computes on exact rational numbers instead of floats, and uses a tolerance that scales with the size and rounding of the statement rather than a fixed amount. A thousands-rounded, correct balance sheet passes; one off by a real amount fails, with an exact reported difference. This also makes the check deterministic (the same numbers run twice produce the identical result), which is what lets a figure be called reproducible rather than just plausible.

When it can’t tell, it says so: it doesn’t guess

Group filings often present figures at more than one scope (parent-only and consolidated), and mixing them produces a confidently wrong answer that looks like a right one. If the figure a check needs at one scope only exists at the other, the check is marked unresolved rather than filled in with whatever’s available. A wrong-but-confident verdict is worse than an honest “couldn’t check this,” because it gets cited as proof of correctness while being false.

The same discipline shows up in how definitions are chosen: equity in the balance-sheet check includes minority interests, because leaving them out would fail every group filing with a partly-owned subsidiary that’s actually correct, and the cash flow check includes the effect of exchange-rate movements on cash, because dropping it breaks the tie on any company operating in more than one currency: choices that match how IFRS numbers actually work, not a textbook simplification.

Every check comes with a paper trail

Whether an identity passes or fails, the result carries the trail back to source: each figure resolves to the exact page it came from. For a passing check, that’s a formality. For a failure, it’s the whole point: “the balance sheet doesn’t tie” isn’t useful on its own, but “the balance sheet doesn’t tie, and here are the four figures with their source pages” is something a reviewer can act on. This is the same audit log discipline that runs through every material action in the platform, applied to the arithmetic.

Who this is for

If you’ve had an AI tool hand you a confident number with no way to check it, or watched a spreadsheet “balance” only because someone plugged a difference into a suspense line, verification is built to stop that quietly happening. It’s what lets a CFO put a figure in front of a board, a lender, or an auditor and say not just “this is our number” but “here is why it has to be right.” Read about the canonical atom store it draws from, see the pillar overview for the rest of the platform, or book a demo to see a deliberately broken filing get caught live.