How we pick
Each day a model puts the stories that cleared a first filter in order, and fixed rules check that order. This page says what those steps are and shows figures counted from the last seven editions.
- Status
- In force
- Operator
- Bruno Pierri
- Contact
- Feedback page
The figures
No daily edition has been stored yet, so the figures that depend on one are not available.
- Stories reviewed a day
- Not available yet
Stories that cleared the first filter and went to the ranking, on average. The model does not see the rest.
- Publishers per Top story
- Not available yet
On average. An outlet, company or account counts once, however many of its channels reported the story.
- Top stories led by a primary source
- Not available yet
The report shown for the story comes from an official first-party source, or from the party the story is about.
- Single-source stories
- Not available yet
Entries in these editions that one publisher alone reported. The rule holds each back from the first three places when another story can take that place.
- Editorial sources
- 490
Sources switched on now whose reports can be picked.
- Corrections, last 30 days
- 0
Entries published in the corrections log.
- Editor agreement
- Not enough reviewed days yet
The share of reviewed Top stories the editor kept as ranked. It is shown once a weekly review exists.
1. The daily edition
Each day has one edition. It covers the 24 hours that end at 1:00 AM ET. While the day runs, the ranking is made again when the stories change, at most once an hour. Shortly after 1:00 AM ET the last ranking is stored, and the daily report is built from it.
2. Which stories are considered
A story goes to the ranking when it has at least one public report in the edition's 24 hours and one of these is true.
- Its best report scores 50 or more out of 100. A model gives each report that score.
- 2 or more independent outlets reported it.
- A page we watch for changes confirmed a change.
- The editor pinned it or promoted it.
Sources that we use only to measure buzz never put a story forward. At most 150 stories go forward in a day, and when more qualify the most widely reported stay. A story that an earlier daily already carried comes back only for a new fact.
3. The ranking
A language model puts the stories in order. It does this 3 times. Each time it sees the stories in a different random order, so no story gains from where it was listed. It sees each story's title, its lead headline, a summary and facts about its coverage, such as how many reports and publishers it has and whether a primary source is among them. It does not see the score. Each pass returns its best 30 stories. We combine the 3 lists by points, where a higher place earns more points.
When 2 or more of the 3 passes say that stories are the same event, they become one entry.
4. Rules that check the order
These rules run in code after the model. The model cannot change them.
- The editor can pin a story to a place, promote it, demote it or remove it. These decisions win over the model's order.
- A story that only one publisher reported, where that publisher is not the party the story is about, is held back from the first 3 places. It takes one only when no other story is left to fill it.
- A vendor's customer story and a compatibility update cannot be a Top story.
- A story cannot be a Top story when more than half of its reports are filed as not AI news.
- The first 10 stories are Top stories. The next 20 are Notable. Notable holds at least 4 research stories and 4 builder stories when the day has that many.
- Between two rankings, a Top story that would move 2 places or fewer stays where it was.
A primary source is an official first-party source, such as a company's own site or blog, or an account of the party the story is about. A publisher is one outlet, company or account, however many of its channels reported the story.
5. The reasons on each story
Each story shows up to three short reasons. Code writes them from facts, not a model. Examples are how many independent outlets reported the story, that the party it is about reported it, how many of the passes ranked it, and that it has a new development since an earlier daily.
6. Mistakes
Models write the summaries and make the ranking, so both can be wrong. When we find a mistake, we add an entry to the corrections log. An entry says what was wrong and what changed. A published entry is not edited, and a fix to it is a new entry.