Apolitics Data
Two records are joined when they look like one company. This page counts how often that is wrong, using something we do not write ourselves: the registration numbers the publishers print. Counted on 2026-10-24.
The usual accuracy figure comes from reviewers' decisions. These are not a reviewer's
| Decided | 16,192 |
| No reviewer at all: the scheduled pass, which joins and never refuses | 10,477 |
| Through an account, in a minute holding more than five decisions | 5,698 |
| At a pace a person can read at | 17 |
17 decisions were taken at a human pace: 14 to join two records and 3 to keep them apart. The rest is the pass approving its own suggestions, including the runs started from an account, which carry that account’s name. On the separate queue for people, 3 of 9,229 decided carry a reviewer. A pass measured against its own approvals reports itself perfect, so the rest of this page measures it against the numbers instead.
One company has one registration number in one register, so a record holding two is two companies
| Records | Hold a number | Hold two | Share |
|---|---|---|---|
| never mergednothing was ever joined into this record | 39,248 | 158 | 0.40% |
| merged unattendedjoined by the scheduled pass, on an equal name | 3,251 | 57 | 1.75% |
| merged through an accountjoined by a run or a click through a signed-in account | 693 | 4 | 0.58% |
A record nothing ever joined holds two numbers of one scheme 0.40% of the time, because publishers themselves print a related company’s number on an entry. A record the scheduled pass joined does so 1.75% of the time. The difference between those two figures is what merging on a name costs.
A merge can only be contradicted by numbers when the record it removed carried one
| Merges by the scheduled pass, with a receipt | 10,186 |
| The record removed carried a number at all | 441 |
| Of those, contradicted by the numbers afterwards | 37 |
4.33% of those merges can be checked this way. The rest joined records that carry no number, and nothing on file confirms or denies them: they rest on a name and on nothing else. That is the limit of this measurement, stated rather than divided away.
Pairs waiting for a decision, and the ones the numbers already settle
| Pairs waiting | 7,898 |
| Their records are on file for numbers that disagree | 248 |
| Of those, pairs the scheduled pass would otherwise take | 42 |
A pair whose two records are on file for different numbers of one scheme in one register is never joined automatically, whatever their names say. The numbers are read from the entries’ own claims as well as from the records, because the number that tells two same-named companies apart often reaches the record only later, when the stored entry is read again.
A Russian nine-digit code is excluded throughout: it names the tax office, not the company, and unrelated companies share one. Two values are only read as a contradiction when they have the same shape in the same register, because one type holds several numbering schemes — a thirteen-digit OGRN and a ten-digit INN are one company’s two codes, not two companies.