Deterministic Calculation

A deterministic calculation always produces the same output from the same inputs, every time, computed by an engine following fixed rules, not estimated, predicted, or guessed by a language model.

Why language models shouldn’t do the arithmetic

A language model predicts the next token. When it produces a sum, it is producing the tokens that most plausibly follow the ones before, which for small, common arithmetic usually lands on the right answer and for long rollups, mixed signs, or unusual magnitudes often doesn’t. The failure is silent: the output is formatted like a correct total, and the same prompt run twice can give two different answers.

Finance is full of problems that have exactly one right answer. Sums, margins, ratios, currency translation, allocations, eliminations, and period rollups are all deterministic problems with defined rules. There is no upside to estimating them and a large downside to getting them subtly wrong, because a small arithmetic error in a driver propagates through every dependent figure in the model.

What determinism buys you

Reproducibility, first. If an auditor reruns the calculation in March on the January inputs, they get the January number. That single property is what makes a figure testable at all, and it’s the reason model risk frameworks are built around validation and recomputation.

It also makes debugging tractable. When a deterministic figure is wrong, the error is in the inputs or the rule, and both are inspectable. When a generated figure is wrong, there is nothing to inspect: the same question asked again may or may not reproduce it.

And it bounds the AI’s role usefully. The model is good at reading documents, mapping messy accounts, explaining a variance, and drafting commentary. It is not good at being a calculator. Splitting the work along that line keeps both halves doing what they’re reliable at.

Where the boundary sits

Determinism applies to the calculation, not to the forecast. A projection rests on assumptions that are genuinely uncertain, and a scenario is a set of assumptions someone chose. What determinism guarantees is that a given set of assumptions produces the same output every run, so a scenario comparison measures the assumptions rather than the noise in the engine.

Rexfin runs those calculations through a calculation engine and hands the AI the result, not the arithmetic; the model reasons over verified numbers instead of re-deriving them, and every figure keeps the citation back to its source. See driver-based forecasting done right.