The two sources agreed on the facts. They disagreed on almost every word.

1. What agreed, field by field

Seven fields per company, 17 companies, under rules written down before any data was collected. One company is a branch and carries no registered capital, so that cell is marked not applicable. That gives 118 comparable values in total.

The three registered-capital disagreements are not disagreements about money. In all three the official registry printed an amount with no currency while the database printed the same amount with one. Our rule, fixed in advance, required both amount and currency to match, so these count as disagreements and we report them as such. Decomposed: 13 matched on amount and currency, 3 matched on amount with the currency absent on the registry side, and 0 disagreed on the amount.

We did not relax the rule after seeing the data. Moving a threshold once the numbers are in is how a measurement stops being one.

2. What a buyer would actually see

Every company in this sample is live. Neither source says so in a fixed way.

Only 2 of the 17 matched literally. The other 15 carried different strings for a status that was, in every case, the same. The registry alone used three spellings for one condition.

The practical consequence is narrow. Suppose your process flags a supplier because the database text and the registry text are not identical. That check will fire on most live companies. It will tell you nothing.

Compare the fields that carry a value instead: the 18-character code, the registered Chinese name, the amount, and the year. Those are values. A status label is not.

3. Method, and what this study cannot say

• Frame: 20 companies were drawn with a fixed seed from the 46 in our registration-field study. 17 were reached before collection stopped, We report 17 and did not fill the gap from elsewhere. Those 46 are themselves the subset that resolved to exactly one candidate by code, so the frame leans toward well-registered entities.
• Registry side: each company opened by its 18-character code in the official National Enterprise Credit Information Publicity System from a mainland China connection, one at a time. Every single query required a human-solved CAPTCHA. Nothing here was scraped, and no CAPTCHA was bypassed, which is also why the sample is 17 and not 500.
• Database side: the commercial record captured on 28 August 2026 for the same codes, already published as our registration-field dataset. One database. This says nothing about commercial databases in general.
• Rules fixed before collection. Status was normalised to five classes. Type was compared after removing spaces and normalising full-width parentheses. Capital was compared on amount and currency, establishment on year, legal representative and domicile on presence only. Business scope was compared on length within five percent.
• The eight-day gap did not hide anything. The registry publishes an approval date, and not one of the 17 had been approved after the database reading. Median approval was 449 days before we looked, the most recent 53 days. So the agreement is not an artefact of nothing having happened yet, and the disagreements cannot be explained away as lag.
• Privacy: no natural-person name is stored. The legal representative is recorded as present or absent only. Company names are omitted from the dataset; the 18-character code identifies each row.
• Limits: n is 17, so every percentage here carries a wide interval. Seven fields from one block of one registry page, one commercial source, one day. Fields were read from the rendered page, so transcription error is possible. Nothing here is a statement about any named company, and nothing here is legal advice.

What the branch record showed

One of the 17 was a branch, not a company in its own right. Its official record carries no registered capital at all.

It names a 负责人 (person in charge) where a company names a 法定代表人 (legal representative). It gives a 经营场所 (place of business) where a company gives a 住所 (domicile).

That matters for anyone building a comparison of their own: a check expecting registered capital on every record will read the branch as incomplete data. It is not incomplete. The field does not exist for that entity form, and a branch obligations sit with its parent. Our own comparison marks that cell not applicable, which is why the denominator for capital is 16 and not 17.

Citing this study

Archived independently of this site, with its own DOI: figshare. If this page ever disappears, the measurement does not.

A citation with everything it needs:

Currawong, “The two sources agreed on the facts and disagreed on almost every word”, 17 Chinese companies read from the official registry on 5 September 2026 and compared field by field against a commercial database record of 28 August 2026. https://currawongweb.com/verify/registry-versus-database-study/

You are welcome to quote or reproduce these figures, including commercially, provided the dates on both sides, the sample size and the stated limits travel with them. Machine-readable evidence: gsxt-vs-aggregator-2026-09-05.json, and the 28 August baseline it is compared against, registration-field-meaning-2026-08-28.json.