When a Filing Is Not Citable: The Floor Rexfin Won't Extract Below

Every number Rexfin shows is supposed to be citable: traceable to a specific filing and a specific page. That’s a straightforward promise on a clean digital PDF. It’s a much harder one on a scan: a low-resolution statutory filing, a fax-quality annual report, a page where the layout has degraded enough that even a careful analyst would squint before trusting a figure. Rexfin’s answer to that difficulty isn’t to try harder and hope the read is right. It’s a floor below which a number simply doesn’t become a citable fact, no matter how much a screen wants to show something.

Why one reading isn’t enough

During ingestion, a page can be read by more than one independent extraction engine. Only when at least two of those readings converge on the same value does that value get promoted into the store as a fact with a citation attached. A number read by a single engine (however plausible it looks on the page) doesn’t clear that bar. It doesn’t get shown, and it doesn’t feed reconciliation or arithmetic checks downstream either. It isn’t held in some lesser-confidence tier; it’s dropped.

The reason isn’t caution for its own sake. A lone-engine reading of a printed figure can be right in one place and wrong in another, with no way to tell which from the extraction alone. The same statement line, read by the same single engine across two comparative periods of the same filing, can land on the correct figure in one period and a neighboring line’s value in the next: same process, same apparent confidence, opposite outcome. At the moment of extraction, a correct single-engine read and a wrong one look identical. That’s the argument for the floor: when right and wrong are indistinguishable, the only defensible move is to hold both back until a second, independent reading agrees.

A tie on the page doesn’t clear everything behind it

Even where a filing’s printed subtotals do reconcile (the components add up to what’s printed), that isn’t automatically enough to keep everything related to that total citable. Rexfin’s checks look at what a total is built from, not just whether the sum matches. If the identity behind a reconciled total ran through even one component that never cleared the agreement floor, the citation on that total is withdrawn too, even though the arithmetic checked out. A number can add up correctly on the page and still not be citable, if the trust behind it leans on a value that never earned it.

There’s a matching rule for gaps rather than errors: when a component is missing because the floor held it back, Rexfin doesn’t substitute a default (treating an absent figure as zero, say) just to force a check to resolve. A held-back input makes that check unresolved rather than silently passing or quietly failing on an assumption nobody asked for.

Graded against a hand-keyed answer, not a hopeful adjective

The bar itself isn’t asserted; it’s measured. Extraction is graded against gold data: statement lines and printed subtotals keyed by hand from real filings and treated as the reference answer for what a page actually says, and those results are what earns a value the right to be called citable in the first place. It’s the same grading discipline behind answer quality evaluation: a standard checked against a known answer, not taken on faith.

What you see instead of a wrong number

When a question depends on a figure that never cleared the floor, Rexfin doesn’t paper over the gap with an invented number or a hedge dressed up as an answer. It says, plainly, that there’s no citable evidence for that figure: the same fail-closed posture described in cite or refuse. A refusal here isn’t a bug report. It’s the system doing exactly what it’s supposed to do when a filing didn’t give it enough to work with honestly.

Who this is for

Anyone who has been handed a confident-looking number and later found out the source page was a bad scan nobody double-checked. If your work depends on knowing not just what a filing says but how sure Rexfin is that it read it correctly, this is the layer drawing that line before a figure ever reaches you. See how it sits alongside the rest of the pipeline at the pillar overview, or book a demo to see a genuinely poor-quality scan run through ingestion and watch which numbers come out the other side.