Offer · Companies shipping AI · 2026-10-03

Is your AI right? A three-day check against the facts.

Your AI feature or agent produces answers, extractions, actions. This audit checks every one of them against a source of truth and tells you, row by row, which are right, which are wrong, and which cannot be checked at all.

FIXED PRICE · THREE DAYS · NO MODEL GRADES THE MODEL · STOCKHOLM OR REMOTE · EU DATA ONLY

Try the free check first: the free check marks every sentence an AI wrote against the document it read, no login, ten pages a day. The audit is that check, built around your feature and run by a person.

← Back to the free check

§ 1 · What you get

Three things, all yours to keep

DeliverableWhat it contains
The reportOne verdict per output (confirmed, wrong, or unverifiable) with the evidence beside it: the register record, the arithmetic, or the line in your own records that decides it. Counts, never percentages without the count behind them.
The harnessThe checking code, handed over. Run it again next week, after the next model update, or on the next thousand outputs. It does not depend on me.
The one pageWhat breaks, how often, and the three fixes that would remove most of it. Written for whoever signs off.
§ 2 · How it works

Three days

Day one. Agree the truth. Every check needs something outside the model to check against. We pick it together: a running balance that must add up, an official register, a confirmation your customer received, a human-labelled set you already trust. If no such source exists for your feature, I say so on day one and we stop there, at no cost.

Day two. Build the check. The check is deterministic. A lookup finds the record or it does not; the arithmetic holds or it does not. No second model grades the first. A model’s opinion of another model is not evidence, and the US Patent Office said as much in a disciplinary order this July.

Day three. Run, report, hand over. The full run, the report, the harness, and an hour walking through what it found.

What counts as a source of truth

Arithmetic inside the document itself. An official register. Your own records of what actually happened. A labelled set you have already checked by hand. Another model does not count, however good it is.

§ 3 · Evidence

Three checks already built, all on real data

DomainThe source of truthResult
Bank statementsThe printed running balance; each row’s amount must match the change in balance742 of 742 rows reconciled on a real 16-page statement; a row the parser had silently dropped was found this way
Trademark reports written by an AIThe Swedish trademark register, 611,097 recordsEvery claim in the report checked, with its application number as proof; the AI’s hedged claims sorted into confirmed, partly confirmed, and not found
Actions taken by an AI agentThe bank statement, which the agent cannot editEach claimed payment confirmed, wrong, or missing; a claimed cancellation contradicted by a later charge; five deliberately planted bugs all caught by the tests

These were built as demonstrations, on my own data and on public registers. They are the method; your audit applies it to your feature.

§ 4 · Scope and price

One feature, one source of truth, fixed price

The audit covers one AI feature or agent, up to 500 outputs, against one source of truth. Larger scopes, more sources, or a standing monthly check are quoted separately after the first audit.

Price

24 000 SEK excluding VAT, fixed. Half on start, half on delivery. If day one shows there is no usable source of truth for your feature, you pay nothing.

Your data stays in the EU. Nothing is used to train anything. The harness runs on your side if you prefer.

§ 5 · Who

Alim Polat, PhD

Former data scientist at a Nordic bank, where a model’s numbers had to be defended to an auditor. PhD in electronic engineering. Based in Stockholm; happy to work remotely across the EU.

Write to me on LinkedIn or at the address on my profile. First reply within one working day.