Holofin

Reading a financial PDF is a specialist job, and it should stay one. Native or scanned, multi-column, tables that wrap across pages, a watermark printed over the holder's name: these are the documents that break general-purpose OCR. Holofin is built only for them. Forest is built for what happens to the answer: which case it lands in, who is allowed to act on it, and what a regulator sees a year later.

An extracted figure is worth nothing until someone is allowed to act on it. Holofin returns each value with its bounding box and the verdict of every validation rule. Forest gives that value a case, an owner, an approval limit and a log line, and applies the same limits whether the reviewer is a person or an AI agent. Neither side has to build the other's half badly.

A reviewer should open one page, not the whole file. When a statement fails to close, Holofin names the rule, the page and the amount. Forest routes on that failed rule instead of on a status, so the right person reads one line, and the correction they make is recorded against the figure. The trail runs from the pixel on the page to the decision, and it still holds when someone asks about it two years later.