How we check
Every night the atlas reads the AI news, works out what actually happened, and checks each claim against the pages it came from before anything is published. This page explains how, so you can judge it for yourself.
The atlas holds 1,128 launches and papers since January 2023 and 134 moves of researchers and executives between labs. Daily briefs started on 28 September 2026, and weekly briefs go back to the week of 21 September 2026. It was last updated on Tuesday, 6 October 2026.
Where the news comes from
Each night the atlas collects new items from 52 sources, which are 16 lab blogs, 10 news sites, 9 newsletters and analysis sites, 4 image, video and audio labs, 4 product changelogs, 4 Chinese labs and outlets, 2 model registries such as Hugging Face, a paper feed, a feed of talent moves and a community forum. A post on social media can point to a story, but the record is built from the pages behind it, the lab's own wherever there is one. A cheap model sorts the items, groups the ones about the same event and keeps the ones that matter, usually 10 to 40 out of several hundred.
How a story becomes a record
For each story worth recording, a researcher model (Claude Sonnet 5.5) opens the primary source, such as the lab's announcement, model card, paper or changelog. It writes the record from what those pages say and nothing else, covering the date, what changed, each figure and who measured it, the price and who can use it. For the bigger stories a second model (Gemini 3.8 Flash) researches the same story on its own, so a mistake by one shows up as a disagreement.
Then each record is checked. A program matches the name, the date and every number against the pages the record cites, and a stronger model (Claude Opus 5.5) reads those pages, settles any disagreement between the researchers from the evidence, corrects what is wrong and marks what it cannot confirm. Only checked records go into the atlas. Nothing is written from a model's memory, so if a page does not say it, the record does not say it either.
What checked means here
- Checked (1,050). The sources support the name, the date, the headline and every figure.
- Corrected (25). Something was wrong, the sources showed what is right, and it was fixed.
- Partly confirmed (47). The core is supported but at least one claim is not, and the record says which.
- Not yet confirmed (6). The record is kept and labelled because the sources did not yet support its core claims.
Figures are quoted as the lab or the evaluator published them, along with who measured them, so a company's own benchmark is labelled as the company's. Every record links its sources so you can open them and check.
The daily and weekly briefs
The written briefs are drafted from the checked records only. They are then read again for style and checked claim by claim against those records before they are published. The weekly brief is revised every day until Sunday.
What readers add
When someone asks the atlas a question it cannot answer from its own records, the answer comes from the web and says so. Each night those questions are reviewed, and any that point to an event the atlas is missing go through the same research and checks as a news story. Nothing a reader writes, and nothing a web answer says, enters the atlas unchecked.
If you spot a mistake, use Report a mistake at the end of any page. A person reads every report, and corrections are made against the sources.
What this does not catch
The checks catch most errors, not all of them. A model can misread a table, and a source can be wrong. Spot checks still turn up the occasional wrong detail, which is why every record shows its sources and its check, and why reports are welcome.
Do it yourself, or let your agent ask
None of this is secret. Every source is linked, and you could follow the same steps yourself. If you would rather your AI agent just asks, an atlas key opens everything since 2023 through MCP for $9.99, paid once. The website and the email briefs stay free.