The Canonical Atom Store: One Verified Fact, One Record

Ask a typical FP&A tool where a number came from and you get a shrug dressed up as an answer: “it’s in the model.” Ask which cell it rolled up from, which import touched it last, which spreadsheet someone pasted it out of six months ago, and the trail usually goes cold. That is not a data problem so much as an architecture problem. Most planning tools were built to hold numbers a person typed in. They were never built to hold a fact that has to answer for itself.

Rexfin’s foundation is built the other way around. Every number that enters the system (revenue, a cost line, a balance-sheet total pulled from an annual report) is stored as a single typed record we call an atom. One atom, one fact, one place it lives. Everything downstream (the model grid, the AI’s answers, the verifier that checks your numbers before an export goes out) reads from that same record. There is no second copy to drift out of sync and no shortcut that lets a different part of the system touch the data directly.

What actually gets stored

An atom is not just a number. It carries everything needed to defend that number later: the currency, the reporting scale (units, thousands, millions, the detail that turns an innocent rounding difference into a false alarm if you get it wrong), the accounting framework, whether it’s a parent-company or consolidated figure, the period it belongs to, and the exact document page it was extracted from. It also carries a count of how many independent extraction passes agreed on the value: a number below that threshold is flagged as not yet trustworthy enough to anchor a hard check.

That last point matters more than it sounds. A figure that only one extraction pass produced is not silently treated the same as one that two independent passes confirmed. The system knows the difference, and so does anything that reads the atom afterward.

This is also what makes click-to-source possible: click a cell, and the drawer that opens resolves straight back to the document and page the atom came from, not to a guess.

Why one store, not several

The alternative (letting different parts of a product read and write financial data through their own private paths) is how planning tools end up with duplicate figures nobody can reconcile. Rexfin’s atom store is exposed through one interface, and every other part of the platform (the model, the AI writer, the export gate) reaches it that way. Nothing reads the underlying tables directly. That sounds like plumbing, but it is the reason a fix to the store doesn’t ripple into three inconsistent copies elsewhere, and the reason a security boundary drawn around one company’s data actually holds everywhere.

That boundary is not incidental. Atoms are scoped to the deal or engagement they belong to, and that scoping is enforced at the query itself, not left to whichever screen happens to be asking. A request for one company’s numbers structurally cannot return another’s, even if some other part of the code forgets to double-check.

The store also groups atoms into statements (income statement, balance sheet, cash flow), preserving which line rolls into which subtotal and with what sign. That structure is what lets a total reconcile against its components automatically, which is the same structure why deterministic calculation depends on to compute anything at all.

What this is not

This is not a data warehouse in the BI sense, and it is not trying to be your general ledger. It holds the facts extracted from source documents plus the figures your model derives from them, nothing more. Rows for documents that turned out to have no extractable financial data still exist; they just carry zero facts, which is a legitimate state, not an error.

It is also not magic. A wrongly scanned page or a badly extracted figure can still enter the store: that is exactly why how verification works exists as a separate layer that checks what landed here before anyone trusts it, and why the ingestion pipeline upstream is held to its own standard.

The takeaway

Most finance software treats “where did this number come from” as a nice-to-have. Rexfin treats it as the precondition for everything else: the model, the AI, and the export gate all inherit their trustworthiness from one plain fact: every atom knows exactly which page it came from, and nothing in the platform is allowed to forget that. For the full picture of how that foundation supports the rest of the platform, see the inside-the-rexfin-platform hub.