Ask a What-If Question, Get a Recomputation, Not a Guess

“How sensitive is operating profit to revenue this year?” is the kind of question that’s easy to ask in a chat box and easy to answer badly. An LLM will happily write you a percentage that sounds right. Rexfin answers it a different way: it changes one disclosed number, walks that change through the actual arithmetic of the statement, and reports exactly which lines move and by how much, every one of them still traceable to the filing page it came from.

What actually happens when you ask

Underneath the question, Rexfin resolves what you’re asking about to a specific reported figure (an operating expense line, a revenue total, whatever the question points at) and a delta, either a percentage or an absolute amount. It then perturbs that one number and propagates the change upward through the statement’s own structure: if revenue is a component of gross profit, and gross profit rolls into operating profit, the engine recomputes each of those subtotals using the same arithmetic that produced them from the filing in the first place. Nothing is estimated. Every dependent line is recalculated, not predicted.

The result comes back as a list of every line the change touched, each with its baseline value, its new value, and the size of the move, plus the chain of statement relationships connecting it back to the number you changed. Because the whole thing runs over the same evidence graph the rest of Rexfin cites from, every affected line still carries its own source page. You can trace a moved subtotal back to the printed figure it depends on, the same way you’d trace any other number in the product.

Read-only, and it says so

This is a probe, not a model change. It doesn’t write anything, doesn’t create a scenario, and doesn’t alter your filed numbers: it computes an answer and hands it back. That matters in a live setting: you can run it again, run a different delta, run it in front of an audience, and there’s no risk of it quietly mutating something you’ll have to clean up afterward. Run the same question twice with the same inputs and you get the identical answer both times, because the underlying arithmetic doesn’t drift between runs the way an LLM’s phrasing can.

The engine is also honest about what it won’t do. Perturb a subtotal rather than a base line, and the totals above it move while the components underneath stay exactly where they were: it doesn’t reverse-engineer how a subtotal’s pieces would have to change to produce it, because that’s a different, much less certain kind of question. And if a change would need to cross between a parent company’s numbers and a consolidated figure, it declines to guess which one you meant rather than picking one silently. Every case where it can’t safely propagate a change is labeled with a reason, not swallowed.

What this isn’t

This is a sensitivity check on numbers you’ve already disclosed: “if this reported figure were different, what breaks or reconciles.” It is not a forecasting engine. Projecting revenue forward, setting a growth driver, or running a full multi-period scenario is a separate capability built for that purpose; see driver-based what-if modeling and the three-statement projection engine for the forward-looking version of this question. Keeping the two apart on purpose is what lets this probe stay simple, fast, and provably safe to run: it never has to guess at an assumption, because it isn’t making one.

Why the traceability matters as much as the math

The reason this is worth having at all, rather than just asking a model to eyeball the sensitivity, is the same reason the rest of the product insists on citations: a sensitivity number you can’t trace back to a filing line is a number you can’t defend if someone questions it. Every moved figure here still resolves to a printed page, the same way a chat answer does: see how the trust chain ranks evidence for how that source linkage is maintained across the product, and provenance drawer, click to source for how you actually get from a number on screen to the page it came from.

For the rest of what’s covered in this part of the product, start at Rexfin product tour.