DATAINT

DataInt collects datasets that public bodies already publish, keeps the address each one came from and the script that fetched it, and stores them in one form. On 14 November 2026 the warehouse held 2,136 datasets and 1,358,691,669 records from 179 named publishers, covering 250 countries and territories. Two reference sites built from it are public, Logibook and DataBook. A third site, VulnBook, is being built on a pool of its own, and the data service itself is not open yet.

This paper counts what is there, says how it was counted, and lists what is wrong with it. Most of the datasets are older than their own renewal schedule allows. A quarter do not come from an official body. One in ten carries a licence under which it cannot be sold.

The work began from seven objections to the way data is sold. Each part below can be read against them, and where the warehouse does not yet meet its own objection, part VI says so.

data brokers hide the source · clickbait datasets · stale data, fresh price tag · opaque pricing · no provenance · pay more, get less · coverage gaps hidden

The warehouse is a tree of folders, one folder to a dataset. Each folder holds a metadata file, and the figures in this paper were read from those files.

Tab. IThe warehouse by section

Section	Datasets	Records
International	131	914,232,393
Countries	1,912	438,360,928
Standards	38	4,600,865
European Union	34	1,242,648
United Nations	21	254,835
Datasets 2,136. Records 1,358,691,669.

The count needs a qualification. Only 355 of the datasets were fetched from a publisher. Another 1,769 are cuts of those, one to a country, made so that a country can be read without opening a global file. The 355 hold 928,683,590 records, 68 per cent of the total.

The pool is also concentrated. Four series hold 54 per cent of all records, and ten hold 63 per cent.

Tab. IIThe ten largest series

Series	Publisher	Records
ILOSTAT	International Labour Organization	349,312,121
FAOSTAT	Food and Agriculture Organization	169,997,342
World Inequality Database	World Inequality Lab	140,961,473
V-Dem	V-Dem Institute, Gothenburg	77,309,794
GeoNames	GeoNames.org	36,442,999
WITS tariffs	World Bank and UNCTAD	31,885,372
GRIP global roads	PBL Netherlands	25,758,453
Official development aid	OECD	11,389,196
World Development Indicators	World Bank	9,749,856
Global Health Observatory	World Health Organization	8,375,460
Ten series. Records 861,182,066.

Every one of the 250 country folders holds at least one dataset; the median is 8 and the largest holds 47. In form, 1,539 datasets are JSON, 502 are SQLite, 90 are GeoJSON, 2 are GeoTIFF and 2 are GeoPackage. One dataset states no form.

Tab. IIIThree counts of the same warehouse

Date	Datasets	Records
8 February 2026	147	23,064,796
20 August 2026	2,125	1,333,548,504
14 November 2026	2,136	1,358,691,669
Between the second count and the third: eleven datasets.

A dataset is admitted to the warehouse when four things exist, and is kept out while any of them is missing. Data written from memory is refused, whether by a person or by a language model, and so is data typed into a script.

1. A source address that leads to the data and not to a home page.
2. The publisher’s name, with the date and the version of what was taken.
3. A script that fetches it again from that address.
4. A metadata file in a fixed schema.

A dataset whose source cannot be found is marked as such and left empty. Of the 2,136 datasets, 2,101 are fetched without a person, 19 need one manual step and 16 are fetched by hand. 2,127 are populated; seven are marked not applicable, one is pending and one needs research. Seven files have no source address and eight folders have no script.

Before a dataset is used it passes four stations: it is scanned, validated, imported and activated.

Two sites are built from the warehouse. Both read it and neither writes to it. On both, the figures are put together by fixed code; what language models do is said at the end of this part.

Logibook is a reference for trade and logistics. It has 58,741 pages in each of 25 languages, and draws on 101 datasets of the warehouse.

Tab. IVLogibook by kind of page

Kind	Entries
Locations	32,695
Commodity codes (HS)	6,939
Airports	6,385
Trade zones	5,290
Ports	3,316
Dangerous goods	2,347
Airlines	1,106
Export controls	382
Countries	250
Incoterms and reference	20
Entries 58,730. With the fixed pages, 58,741.

DataBook is a reference by country. Its build holds 244 countries, each described in 15 sections and 25 languages, and ranks the countries on 51 indicators. The sections are population, economy, labour, education, health, governance, parliament, defense, infrastructure, environment, geography, identity, compliance, space and timeline.

Labels and descriptions are translated into the other 24 languages by language models, and DataBook’s prose passages are written by one. A reader who needs a figure should take the figure; a reader who needs its meaning should read the English.

VulnBook is the third site. It does not read the warehouse described above. It keeps a pool of its own, of public vulnerability records, and joins them around the product: a reader who runs one piece of software is shown what that software contains and what is known against each part.

It is not public. The project measured itself on 7 November 2026.

Tab. VVulnBook as its own project measured it

On 7 November 2026	Count
Sources in the catalogue	62
Sources that had reached the normalised layer	24
Sources that had reached a page	0
Observations kept, in millions	20.6
Vulnerabilities in the normalised layer	687,725
Sources 62. On a page, none.

On that date the relation between products and versions, which is the point of the product, existed in no layer. It is being built now. A first vulnerability page passed review on 13 November and runs at a test address; no page is at the public one.

Three of its rules answer part I directly. An observation is never overwritten: a correction is added beside what it corrects. Every published claim carries its source and the time it was seen. The site gives no risk score of its own, and an empty result reads “no known record”, not “no vulnerability”.

The destination of all this work is the data service itself: access to the warehouse by program, with the source of every row attached. It is not open. The sites are waypoints on the way to it, and not the thing itself.

Four things limit what this warehouse can be trusted for today: the age of the data, where it comes from, what its licences allow, and how it was counted. They are taken in that order.

Age. Each dataset states how often its source is renewed. Measured against that statement on the day of the count, 1,234 datasets were overdue. Every dataset that promises a daily, weekly or monthly renewal is late.

Tab. VIStated renewal against the last one made

Renewed	Datasets	Overdue
Daily	13	13
Weekly	11	11
Monthly	717	717
Quarterly	259	253
Twice a year	244	240
Yearly	343	0
Every two years	247	0
Datasets 1,834. Overdue 1,234.

The newest dataset was last renewed 82 days before the count, and the median age is 212 days. No file records a renewal in July, September, October or November. This is the third objection of part I turned on its author.

Tab. VIIMonth of the last renewal

2026	Datasets
February	206
March	669
April	902
May	276
June	57
July	0
August	24
Datasets 2,134. One more is dated 2017 and one is undated.

Origin. 1,608 datasets, 75 per cent, come from an official body: 1,466 from an international one, 112 from a government, 16 from the European Union, 12 from an industry standards body and 2 from an official non-governmental one. 515 come from a public reference source or a non-governmental organisation, 4 from an open community and 1 from a research group. The project’s own introduction calls the sources entirely official. The metadata does not support that sentence.

Licence. 917 datasets are under CC BY 4.0, 242 under CC BY 3.0 IGO and 109 under a government open licence. 213 carry a licence that forbids commercial use or is proprietary, and a service that sells access cannot include them. 624 files say only “open”, which is a description and not a licence.

Counting. The record totals are the sums of what the metadata files report. The share of derived datasets makes the number 2,136 look larger than the number of things fetched, which is 355. The public landing page of DataInt still shows an earlier count of 2,055 datasets.