Every verdict in the release
Most of what could be checked held up.
Each square is one claim the record could settle. 2,164 of 2,653 were true or mostly true; 62 resolved false.
- 62 resolved false
- 183 mostly false
- 244 mixed
- 1191 mostly true
- 973 resolved true
Hover or tab to a square to see whose claim it is; that person's other verdicts light up. Select it to read those claims.
28,548claims extracted from public output
10,415checkable: a fact, a promise, guidance or a prediction
3,777researched against the public record
2,653resolved, 70% of those researched
47%of verdicts carry a second judge's dissent
Every score on this page is for one kind of claim. Predictions are included and scored on their own; they are never averaged in with facts in a chart.
- Factual claims
- statements about the present or the past, checked against the record: did they hold up? 2,177 resolved.
- Predictions
- statements about events outside the person's control, checked once they fell due: did they come true? 340 resolved.
- Promises
- commitments to the person's or organisation's own future action, with something to deliver: was it delivered? 128 resolved.
- Guidance
- a company's own figures for its results over a period, checked against what it then reported. 8 resolved.
Person by person
Most false verdicts are predictions.
40 of the 62 claims resolved false were predictions, though predictions are 340 of the 2,653 claims the record could settle.
Showing: all kinds of claim together, people in alphabetical order.
Each bar is one person's resolved claims, centred on the mixed ones. With all kinds of claim together the bars are not ranked: they mix facts, predictions and promises in different proportions for different people. Choose a kind of claim to put people in order of their score for it. True or mostly true means the claim matched the record the method found; false or mostly false means it did not; mixed means the record supported part of it. Nothing here is a judgement of the person. Select a part of a bar to read those claims, or a name for all of that person's.
The scores and their ranges
Scored separately, most scores still overlap.
495 of 527
pairs of people compared on the same kind of claim have overlapping 80% ranges. A score is a range, and two people a few points apart are not reliably in that order.
One panel per kind of claim, each ranked on its own. Only scores at medium or high confidence are ranked; the rest are named under each panel, and every score is in the table below. The dot is the score, the line its 80% range; a filled dot is high confidence, a ring medium. The shaded band is the range of the highlighted person; everyone it touches could plausibly hold the same score. 8 guidance claims were resolved in all.
A prediction score counts how many predictions came true. It does not yet allow for how likely each was when it was made, so an easy prediction counts as much as a bold one.
Predictions against facts
Predictions hold up less well than factual claims.
20 of 21
people with both scores do worse on predictions than on factual claims, by 15 points at the median. Each prediction score rests on 5 to 44 claims, so any one gap is uncertain; the pattern across people is the finding.
Only people with enough resolved predictions for a score are shown (21 of 33). Promises were scored for 9 people; every score is in the table below. Each line compares one person with themselves, not people with each other. A dashed ring is a prediction score at low confidence.
Where the record runs out
About 70 in 100 researched claims could be settled.
Up to 60 checkable claims per person were chosen to cover their main topics and years on record, then researched. The rest of the researched claims are unresolvable: the record was too thin to decide, which is not the same as false.
Posts against pages
Posts on X against pages.
873 checkable claims came from posts on X, read through the X API; the rest from interviews, talks, articles, filings and transcripts on the open web. Of the claims that could be settled, 86% of posts and 81% of pages held up. 31% of researched posts and 30% of researched pages were unresolvable.
Read with care
What the numbers do not say.
A second judge agreeing is consistency, not accuracy. A model judge (Claude Opus) gives the verdict. A second judge from another model family (OpenAI GPT) reads the same claim and evidence independently; where it disagrees, the claim is flagged. In this release 47% of verdicts carry the flag. The method's agreement with published outcomes is measured separately on the test results page, and it is not a general measure of accuracy.
Unresolvable is not false. 1,124 researched claims ended without a verdict because the record was too thin. They are left out of every score.
Party-political claims are scored in reports made from 29 October 2026. 530 claims are tagged political in the data; 1 of them, in reports made before that date, were left out of scores, as the method then required.
Public output only. Claims come from what each person said in public in their professional capacity. Nothing about private life, and no inferred motives. Each person can respond through the reply route; replies are attached to the report.
Verdicts are made automatically by the published method, versions 0.1.0+03d23c0e, 0.1.0+28da144b, 0.1.0+3e03e2df, 0.1.0+3f84d14d, 0.1.0+7831dd9d, 0.1.0+8bea457b, 0.1.0+c876030c, and are not individually reviewed by a person. A verdict is a statement about the record the method found on the day, with links to it, so anyone can check it.
Open data
Get the data.
Release v1.1.2 is three tables, CSV and JSON Lines, under CC BY 4.0: one row per person, one per score, one per claim with its quote, source URL, verdict, confidence, the judges' flag and the evidence count.
Read every claim in your browser at /explore. Search the quotes, filter by person, kind of claim and verdict, and open any claim for its evidence.
The code and prompts that made it are not published; the method is. One person, no key, from the command line or from a model's tool call:
curl "https://epistacklabs.com/v1/lookup?name=Jensen%20Huang"The developers page describes the lookup, the request queue for names not yet in the database, and the full API.
Every figure
The table.
Every figure in the charts, as a table. Scores are followed by the number of resolved claims in brackets. Each kind of claim has its own columns; the combined score, last, mixes them in the proportions given beside it. Built 2026-11-10 from release v1.1.2.