Who catches the error

Who catches the error

1. A contract arrives in different shapes: a crooked scan with a seal, a Word file, or a neat PDF with a table. Each holds values that money and deadlines depend on: the company code, the account number, the amount, the date. Three such contracts show the ways to check a value in a document and which error each way catches. They also show why a neat file deserves no more trust than a scan.

2. How a document looks depends on how it was made, not on whether its data are right. A PDF is usually saved from the same Word file, and a scan is a printed sheet that was photographed. In all three, someone typed the code, the amount and the date on a keyboard or carried them over from an earlier contract. A typing error passes into any format unchanged, so the check is the same for all three.

3. The first check needs nothing but the value itself. The EDRPOU code, the number under which a company is entered in the state register, has eight digits, and the last one is a check digit: it is computed from the first seven so that a typing error shows at once. Each of the seven digits is multiplied by its weight, the products are added, and the sum is divided by 11. For code 31316718 the sum is 96, the remainder 8, and the eighth digit is 8 too.

4. In the scanned contract the same code is written as 31361718: two neighboring digits have swapped places. The products change, the sum is now 91, the remainder 3, while the eighth digit is still 8. Such a mismatch means the code most likely contains an error, and it gets checked against the register. The formula almost always catches a single wrong digit the same way. A remainder of 10 has its own rule: the calculation is repeated with other weights.

5. This kind of check is called algorithmic: it works wherever a value has a formula. An IBAN carries two check digits of its own, and the one in the scan fails them. In the Word contract the product of 15 and 12,400 is written as 168,000 instead of 186,000, and the PDF has the date April 31, which is not in the calendar. Yet code 47645196 from the same PDF passes the formula: it checks how a value is written, not whether the company exists.

6. Only the register answers whether a company exists. The EDR, Ukraine's state register of legal entities, has an open search that shows a company's name, address and director by its code. For code 47645196 from the neat PDF the register finds no entry. That does not yet mean the company does not exist: the error could be in the code itself, so the search is repeated by name. The buyer does have an entry, but the contract says building 6-B and the register says 6-K.

7. What remains are errors that neither the formula nor the register sees: each value is plausible on its own, but they contradict each other. Cross-reading finds them: a person compares different places in one document. The contract is dated May 21, while the annex refers to a contract of May 12. In the preamble the director's initials are O. S., under the signature they are O. V. The warranty in clause 6.1 is 12 months, in the annex 24.

8. An AI assistant, a program that answers a request with text, speeds up all three checks, but the result depends on the wording. To a general “check the contract” it answers confidently and may write “no errors found” without comparing the dates. To a precise “compare the dates in the contract and annex 1” it quotes both dates and shows the mismatch. A confident tone says nothing about how much it actually checked.

9. In each of the three checks the assistant has its own share of the work and its own limit. For formulas it copies codes, amounts and dates into a table, but the computing is better left to a formula, since the model makes arithmetic errors. It searches the register only when the product gives it that access; without access the answer is a guess. In reading it finds the paired places, and a person decides which of the two values is right.

10. The three checks catch different errors, so none replaces the others. You start with formulas, because they are fast and cost nothing, then comes the register query, and the slowest part, reading the links between clauses, stays with a person. A scan, a Word file and a neat PDF go along this path the same way. The neat PDF in this example turned out to have both a date that is not in the calendar and a code that is not in the register.