How Deterministic Math Keeps FactoryOS Honest
Ask a language model to multiply two numbers and it does not calculate; it generates what a right answer looks like. That is fine for prose and disqualifying for your books.
FactoryOS resolves this the only way that actually holds: the AI is never allowed to do arithmetic at all. Every calculation is handed to a deterministic math engine, and every figure that comes back can be audited line by line.
Why AI Makes Up Numbers
AI makes up numbers because generating plausible text is the entire mechanism, and a plausible number is not a computed one. The industry word is hallucination, and for arithmetic it means "looks right" and "is right" are two different values the model cannot tell apart.
No amount of model improvement removes the problem, because guessing well is still guessing. The fix is architectural, not cosmetic.
The Model Never Touches Arithmetic
In FactoryOS, the model's job ends at setting up the calculation: which formula, which inputs, which data. Its standing instruction is literal -- never do the arithmetic yourself.
The math itself runs in plain deterministic code, with no model inside the engine and no network egress. The AI chooses the question; a calculator answers it.
Same Input Same Output
Deterministic means the same input produces the same output, every single time. Run the calculation today, next quarter, or in front of an auditor, and the figures match to the digit.
That is the property a spreadsheet has and a language model never will. FactoryOS simply refuses to let the part that guesses anywhere near the part that counts.
Every Answer Shows Its Work
Every calculation renders as a computation card: the headline answer prominent at the top, the original formula it was asked to solve, the inputs used, and the worked steps underneath. You see the formula with your values substituted in, the reduction line by line, and the result.
Results are formatted for their meaning as well -- currency renders as currency, percentages as percentages. The answer reads the way your controller would write it.
Audit Any Number in Seconds
Trust, but verify is the design, not a slogan. The main answer is for reading; the work laid out below it is for checking, and the source of the data rides along with the result.
When a calculation runs over your real records -- a column of invoices, a quarter of latency logs -- the engine pulls the values directly, and the raw numbers never pass through the model at all. What gets computed is your actual data, verbatim, not the model's recollection of it.
It Refuses Rather Than Guesses
When the engine cannot compute something, it says so in plain language instead of producing a number anyway. An unknown formula gets a "did you mean" suggestion; division by zero gets an explanation, not an invention.
That refusal is the point. A system that can say "this is undefined" is a system whose answers mean something when it does answer.
Formulas Layered by Discipline
The engine is built in layers: a full scientific calculator at the base, statistics in the middle, and named industry formulas -- finance, billing, budgeting, marketing, unit conversions -- on top. Dozens of named formulas across seven disciplines today, with new disciplines added as whole layers, and the catalog growing toward the professions that must verify their math.
Each named formula is verified in code once, then reused forever -- net present value, loan payments, percentiles, regression -- rather than reconstructed from a model's memory each time. The formula you call Tuesday is byte-for-byte the formula you called Monday.
What This Means for Your Books
It means the question that kept AI away from your numbers has an answer you can inspect. The reasoning is prepared for a human to approve, the arithmetic is exact, and the work is on the screen.
Numbers are where trust in AI is won or lost, which is why an AI you own should be strongest exactly there -- including when the number in question is the return on the AI itself. What is the first calculation your business would want to see done in the open?