Never Let a Language Model Do Your Arithmetic

You stop AI producing wrong numbers by refusing to let a language model produce numbers at all. A language model predicts the next plausible token, so it is built to generate text that reads correctly, not a figure that reconciles. In Mickai, deterministic engines calculate every value from the source data, and the model is confined to the language around those values. The same engine version on the same inputs returns the same result and the same digest, which is what a reviewer needs.

• A language model predicts plausible tokens, and fluency is not arithmetic.
• Deterministic engines produce every figure, so the same engine version on the same inputs returns the same result.
• Each result carries a SHA-256 digest, so a number can be matched byte for byte across runs.
• Every consequential action is sealed into the Open Audit Record, an append-only, hash-chained, tamper-evident log.
• Entries are signed with FIPS 204 ML-DSA and verified cold and offline with the operator public key alone.
• The model writes the sentence around the number. It never writes the number.

Why does a language model get numbers wrong?

A language model is a probability engine over text: given the tokens so far, it emits the token most likely to follow. It excels at phrasing, summarising and drafting, and holds no notion of a balance, a carry, a rounding convention or a control total. When it prints a figure, it is producing the shape of a number that fits the sentence.

Two failure modes follow. A wrong figure reads as confidently as a right one, because nothing in the mechanism distinguishes them. The output is also unstable: ask twice and the wording moves, sometimes taking the value with it. Nobody can sign off a figure that will not repeat.

How do you keep the model out of the calculation?

The split is architectural, not a matter of prompting discipline. A deterministic engine reads the source records, applies the stated rule, and returns a value with the inputs it consumed and a SHA-256 digest of the result. No sampling and no temperature exist in that path. The model sits beside that engine and never inside it: it interprets the request, selects the engine and the parameters, and writes the sentence around the returned figure.

Determinism here is a claim made against a version: the same engine version, applying the same rule to the same inputs, returns the same value and the same digest, today or long after the year end. Change the rule or upgrade the engine and the figure stays reproducible against the version that produced it.

How do you prove a figure has not changed since sign-off?

Every consequential action is sealed into the Open Audit Record. Each entry carries an index, a timestamp, the actor, a typed action, the target and a SHA-256 of its payload, and entries are hash-chained, so altering, removing or backdating an entry breaks the chain from that point on. The record is append-only and tamper-evident, which means it does not prevent an attempt, it makes the attempt detectable to anyone who verifies.

Entries are signed with FIPS 204 ML-DSA, the post-quantum signature standard, and anyone holding the operator public key can verify the record cold and offline. That property rests on custody of the signing key: a party holding it could rewrite and re-sign, which is why signed checkpoints written off the box and anti-rollback matter, making truncation detectable too.

What stops the system acting on a number by itself?

Consequential actions are staged: the system prepares the action, shows the figure, the inputs and the rule applied, then waits for a named person to approve or refuse. The approval or refusal is sealed into the record, so who accepted a number, on what evidence and when, has one answer.

That matters most where a figure triggers a payment, a filing, a disclosure or a covenant test. Staging keeps the accountable person where accountability already sits.

Why does determinism matter more in finance than anywhere else?

Regulated reporting has a property most computing does not: the answer must be defensible long afterwards, by a different person, from the same evidence. Determinism turns a claim into a test.

It also changes the shape of a review. Rather than re-performing the work, a reviewer confirms the engine version, the inputs and the rule were the intended ones, then verifies the digest. Scrutiny moves to the inputs and the policy, where judgement belongs.

What is a language model still the right tool for?

Everywhere language is the work: drafting the narrative in a set of accounts, explaining a variance, turning a vague request into the right query, reading a long contract and pointing at the clause that changes a treatment. None of that needs the model to invent a value.

Mickai runs on hardware you own, over loopback only, with no outbound path by default, so ledgers and working papers stay on the machines you control. Mickai LTD owns 104 filed UK patent applications carrying approximately 2,340 claims, an estate that spans the architecture and is working towards examination.

Frequently asked questions

Can I just tell the model to use a calculator?

Tool calling helps, and it is not a guarantee. The model still decides whether to call the tool, which tool, and what to pass it, and it can still paraphrase the answer. Here the engine produces the value and its digest, and the model may not recompute the figure it receives.

How do I know the figure in the report is the figure the engine produced?

The figure carries a digest, and that digest is sealed in the Open Audit Record. Recompute the SHA-256 of the result and compare it with the sealed entry. If a value was edited between calculation and publication, the comparison fails and the chain indicates where.

Does this work in an environment with no internet access?

Yes. Mickai runs on your own hardware, over loopback only, with no outbound path by default. Verifying the record needs the operator public key and nothing else, so an auditor with no network can confirm that a figure and its approval have not moved.

Is deterministic AI just a spreadsheet with better marketing?

Spreadsheets are deterministic, and mainstream ones do keep version history and change tracking. What is distinctive here is that the rule, the inputs, the result and the approval are sealed together as one entry in an append-only, hash-chained record, signed with a post-quantum signature and verifiable offline with the operator public key alone.

How do I see this working on my own reporting?

We run a selective closed beta at mickai.co.uk/beta, and not every applicant is accepted. If your team produces numbers that must reconcile, and be defended long after their author has moved on, bring a reporting process that eats review time and we will walk this split through it.