How good is the data?

Every part of the graph is listed with its measured accuracy, or marked UNMEASURED if nobody has checked it yet.

Checked

84.0%correct

94.7%faithful to the source text

300 facts checked by hand on July 31, 2026, drawn at random from a batch of 3,346 read from English Wikipedia that month. Most of the errors were links to the wrong entity rather than misreadings of the text, which is why the second figure is higher. Linking is being improved.

These numbers are for that batch only, not the whole graph.

Still to check

Most facts read from text

About 2,900 facts have been checked by hand so far. The figures on this page come from the ones drawn at random, which cover a few batches: the July batch and the two below. Every fact read from text also goes through the graph's automatic type and date checks. More samples are next.

A batch with no sample is marked UNMEASURED rather than left blank.

Two more samples, read blind

Facts read from text and linked to an entity, drawn at random and read by a person who could not see the model's score:

11.9% wrong in the September 2026 batch.
13.8% wrong in an earlier sample, covering most of what was read before September.
97.1% correct when Wikidata holds the same fact, against 86.0% when it does not. That agreement now ships as a field on every linked fact read from text.

Fixing wrong links

Most wrong facts read from text are links to the wrong entity with the same name, and fixing them comes first. People read them, and models carry each fix further: in the current release a person ruled 181 wrong, and a model confirmed 219 copies of those same links. A model then removed 2,052 more, keeping the words read from the text but not the wrong link.

Of removals drawn at random, a person reading blind confirmed 98 of 100 as wrong links.

The model's own confidence

The model scores every fact it reads from text, and it is sure of nearly all of them: 90.5% score 0.9 or higher. Checked by hand, the July batch came out 84.0% correct, for that batch only. Most of the false facts found so far scored 0.9 or higher as well, so the score is shown here and not trusted as a filter.

1.0: 42.2%
0.9–1.0: 48.2%
0.8–0.9: 7.6%
0.7–0.8: 1.5%
under 0.7: 0.3%

Facts read from text only, by the model's score. Facts from Wikidata are scored 1.0 or not at all, so they are left out.

Why the two kinds are checked separately

They go wrong in different ways, so they are checked in different ways, and never averaged.

From Wikidata: matched by identifier, not by name: Nothing is matched on spelling, but that does not make it error-free. Wikidata's mistakes come along with it, and importing it has had bugs of its own, such as part-of facts recorded backwards. Those found are being fixed release by release, and the ones in this sample still wrong are left out. No random sample of this part has been checked by hand yet, so its accuracy is UNMEASURED.
Read from text: unsure means no link: When the model can't match a name to an entity, it keeps the name and makes no link. When every such name was checked against the graph's labels, 92.1% matched nothing. That was a check by exact label, on August 26, 2026.

Checked again for this site

The graph on this site is a small sample, and it was checked again before it went up.

156 facts read from text are drawn in it. Each was checked by hand against the text it came from before it went in, and an approval is withdrawn when the fact turns out wrong.
41 facts from Wikidata are left out: recorded backwards, out of date, wrong, or too vague to show. They stay in the graph.
What turns up wrong here goes back to be fixed at the source, and each new release is checked for it before the sample is rebuilt.
The dates and facts in the stories were checked against published sources.

If a wrong fact would cost you something, this is the page to read first. If it is your own subject that needs checking, talk to us about it.

Where each figure on this page comes from

84.0%: correct, on the July 2026 batch: Gold labeled Checked by hand against a labeled sample.A random sample of one batch. It describes that batch and does not carry over to the whole graph.
94.7%: faithful to the source text, on the July 2026 batch: Gold labeled Checked by hand against a labeled sample.Same sample, same limit: it describes that batch only. It counts misreadings of the text; a link to the wrong entity counts against the first figure, not this one.
300: facts checked by hand in that sample: Gold labeled Checked by hand against a labeled sample.Drawn at random within that cohort.
3,346: facts read from text in the July 2026 batch: Release report Computed with each release, from its own quality report.One batch, read from a slice of English Wikipedia that was not chosen at random.
2,900: facts checked by hand so far, about: Gold labeled Checked by hand against a labeled sample.Across every sample checked so far, including ones drawn for training and calibration rather than to measure a batch. Only the random samples give the figures on this page.
181: wrong links a person read and ruled wrong: Build record Measured while building BaseKG, and recorded in its run notes.Read one by one by a person. In the current release.
219: copies of those wrong links, each confirmed by a model: Build record Measured while building BaseKG, and recorded in its run notes.The same wrong link appearing elsewhere, each copy confirmed by a model before it was fixed. In the current release.
2,052: wrong links removed by a model, then audited by a person: Build record Measured while building BaseKG, and recorded in its run notes.The link is removed, not the fact: the words read from the text are kept, unlinked. In the current release.
98 of 100: removals confirmed wrong on a blind read by a person: Gold labeled Checked by hand against a labeled sample.100 removals drawn at random, each read by a person who could not see the model's verdict.
11.9%: wrong, on a blind read of the September 2026 batch's linked facts: Gold labeled Checked by hand against a labeled sample.19 of 160 facts drawn at random from that month's batch of facts read from text, among those linked to an entity, each read blind by a person. Wrong in any way counts. 160 is a small sample, so the true rate could be a few points either way.
13.8%: wrong, on a blind read of most older linked facts read from text: Gold labeled Checked by hand against a labeled sample.22 of 160 facts drawn at random from most of the facts read from text before September 2026 and linked to an entity, each read blind by a person. Wrong in any way counts.
97.1%: correct, when Wikidata holds the same fact: Gold labeled Checked by hand against a labeled sample.34 of 35 hand-checked facts read from text that Wikidata also holds. Facts Wikidata also holds may simply be easier ones, so this predicts, it does not explain.
86.0%: correct, when Wikidata does not hold it: Gold labeled Checked by hand against a labeled sample.The rest of the same hand-checked set.
92.1%: of unlinked names matched no entity in the graph: Recovery audit Measured on the samples where missing links were looked for a second time.Every name the model could not link, checked on August 26, 2026 against the labels of every entity in the graph. A name could still match under another spelling, which was not tested. More names have been left unlinked since.
156: facts read from text drawn in the sample, each checked by hand: This sample Counted in the sample this page is drawn from.Checked by hand, one by one, against the text they were read from. Ones that ended up outside the sample are not counted.
41: facts from Wikidata left out of the sample: This sample Counted in the sample this page is drawn from.Each named by its id: recorded backwards, out of date, wrong, or too vague to show. They stay in the graph; leaving one out of the sample does not delete it.