OK
The line is correct, reconciled to the øre.
Labour market contribution 8 %of 14.859,15 kr.
−1.188,73OK8 % of the base is 1.188,73 kr. That is correct.
From the demo payslip. ATP and income tax (A-skat) are also OK.
//How it works
Language models read the Danish collective agreement and extract the rules. A rule engine with no language model calculates the pay, so the same payslip gives the same answer, every time.
//Three steps
The language model reads. The rule engine calculates. Between them sits the requirement for a source.
Language models extract the rules with their period of validity and conditions. That is reading, and reading is what language models are good at.
A rate only applies with a verbatim quote or a reference to a table cell in the collective agreement.
No language model in the calculation. The sums are checked with mathematical logic, and the same input gives the same answer, byte for byte.
Every answer has a clause, a verbatim quote and a version stamp, so it can be verified. See Security
//Six answers
Three of the examples are from the demo payslip: a synthetic Industry apprentice payslip, step 2, March 2026.
The line is correct, reconciled to the øre.
From the demo payslip. ATP and income tax (A-skat) are also OK.
An error, with the amount in kroner and the clause.
From the demo payslip.
Something a professional should look at.
Typical example.
We're missing one document, and we say which.
From the demo payslip.
The rule isn't in our rule set yet, and we say so.
Typical example.
The line doesn't fall under the collective agreement.
Typical example.
No line is hidden. Every line is either checked, or it says what we're missing.
//The whole case
Upload two years of payslips at once. PayTjek follows the case month by month: step changes, rate adjustments, changes of collective agreement and year-to-date balances. The result is one combined claim.
Only one in four errors can be seen on the payslip alone. Over half require the collective agreement, the law or the contract, and almost one in five requires the previous payslip.
//Why not just a language model?
A language model predicts the most likely next word. It learns patterns, not rules. That works well for text and conversation.
But a pay line has only one correct amount. "About DKK 190" is not good enough.
Language models are pattern machines. Pay calculation needs a rule machine.
A payslip has many lines, and a collective agreement has hundreds of rules. On top of that come the law, local agreements and the contract.
A rule that looks irrelevant can change another. Everything has to be right at once, and small error rates grow quickly.1
The more a language model has to keep track of, the less reliable it becomes. The research below shows this.
Pay calculation is amounts, rates, date ranges and rules that interact. One wrong digit gives a wrong answer.
Language models are built for free text and conversation, not for calculating to the øre.
That is why we don't calculate with a language model. We read with it.
//Sources
Language models make best use of information at the beginning and end of a long text. When the important part is in the middle, accuracy drops markedly, even for models built for long texts.
18 language models tested. The answers became less reliable the longer the input was, even on simple tasks.
Send us the collective agreement, the local agreements and a case in advance. Book 30 minutes. We'll run your own payslip through while you watch.
//Footnote
1 A thought experiment, not a measurement. If a language model has to get 500 rules right at once and gets each one right with 99.5% certainty, the chance of getting all of them right is 0.995500 ≈ 8%. This assumes the errors are independent of each other. Back ↑