How GetYourBrief works

How reports become one event

Each report is turned into a text embedding. Reports whose embeddings are nearly identical (cosine similarity 0.92 or more) are treated as the same report; reports that are close (0.85 or more) are grouped as one news event. An event can therefore carry one report or a hundred. Until October 2026 an event was written up once per desk it touched; since the consolidation of the 17 desk sites into GetYourBrief, each event has one home and the other write-ups point to it.

How summaries are written, and what is checked

For each event a large language model (from October 12, 2026, Xiaomi MiMo V2.6 Pro, with DeepSeek V4.1 Flash when it fails; earlier events, DeepSeek V4 Pro; rewrites for legal safety, DeepSeek V4.1 Flash) writes a headline, a short summary and a list of key facts from the text of the reports. Before anything is published, automated checks run on every write-up: the key facts must exist and number at least three, the headline and summary must be present and within length, at least one person, company or body must be named, and any year later than the current one or earlier than 2020 is flagged as a possible error. A write-up that fails those checks goes to a second model review, which can reject it; a rejected write-up is not published. If that review cannot run, the write-up is held, not published (from October 11, 2026; before that date it was published without the review). Events whose reports carry no usable text (only a headline, or cookie and navigation boilerplate) are skipped rather than written up from headlines.

Every headline, summary sentence and key fact then passes a legal-safety screen before it is published, and the whole archive passed the same screen on October 11, 2026, followed by a check that every displayed statement is supported by the reports listed with the event on its page (a statement they do not support is removed; a headline they do not support makes the event a coverage record). A decision model (TypeSafe Jev) asks five questions of each sentence: does it state that an identifiable person or company did wrong, rather than reporting who alleges it; does it give an opinion or characterisation in the site's own voice; does it address or advise readers; does it name a private individual in a sensitive matter; does it predict a legal, medical or financial outcome as fact. A sentence that raises any of them is rewritten by the language model to state only the verifiable step and who took it, with the accusation marked as alleged, and is screened again; if it still fails after two rewrites it is removed. An event whose headline cannot be made safe is written again from its source reports under the same rules and screened again (the archive's 2026-10-11 pass did this for every event the first screen withheld or left with under two sentences); an event that still cannot be made safe stays on the site as a neutral entry: a label naming the company or body involved and the date, the counts and the links to every report, with no summary of ours. The thresholds are set so that a doubt costs a rewrite, not a publication. The model-written timelines of earlier versions are no longer shown: an event's development is its reports, each at its own time.

Three more checks run on every event. Each report listed under it is read against the event, and a report the clustering attached that is about something else is neither listed, counted nor used for these checks. Each headline, sentence and key fact is read for whether it keeps who did what to whom and the reports' own certainty ("may", "suggests", "alleges"); one that does not is removed, and a headline that does not makes the event a coverage record unless a rewrite passes every check. Each headline is also read for a number, range, office, role or status the reports do not give (a range given as one figure, a planned cost given as a spent one, a task-force seat given as a board seat), with the same outcome. Since October 12, 2026 every headline, sentence and key fact is also read against the listed reports for a figure, role, status or certainty given differently, and for a claim, view or estimate given to someone other than whoever made it. A coverage record's address names no one. It states the facts of its coverage (how many outlets, between which times, who first) and shows the headline an outlet gave it, quoted and attributed to that outlet, when that headline passed our screen; otherwise it is a line listing its outlets and links. A report's own title, under any event, is shown or quoted only when it passes that screen, and never when it concerns an arrest, a charge, a prosecution, a trial, a sentence or a violent crime, or alleges, accuses or reports a probe, a lawsuit, a settlement or a scandal, whoever it names; organisations are not listed beside such reports. An address is read like a headline, and one that states an accusation is replaced by one that names no one; a headline rewritten for these checks gets an address made from the new wording. A report's own title is shown only after the same legal screen has read it and passed it; until then the report is named by its outlet. And the event's address and its reports' titles are read for a private person named in a crime, court, health or similar matter, or anyone under 18: such an event gets an address that names no one, and its reports are listed by outlet and link without their titles. A report's text is used for these checks and for the summary only where the outlet's robots.txt, TDM header or tdmrep.json allows it, and not when our check finds it to be the copy of a news agency that reserves its text (AP, Reuters, AFP, Bloomberg), wherever that copy is carried. The check reads the whole stored text for a dateline, credit, sign-off or copyright line, the agency's own site and web addresses that file a piece under the agency, and copies of such a report (the same headline, or a fifth of its text shared); every report from a site where at least half of the stored reports are agency copy, or that runs agency copy without the credit, is treated as agency copy too. Of a report whose text is not used, only its title is. Agency copy that carries no sign of its source in the text we store can still pass the check; when one is found, it is taken out and its event checked again.

These checks test wording and structure, not truth: no person reads every summary before it appears. That is why each one sits next to its reports, and why a correction, an objection or a takedown request (corrections, people named) removes a sentence, a name or an event from every page within 15 minutes.

At each review the data editor checks a random sample of key facts against the reports they cite. Last check: 2026-10-11, 11 events and their 65 key facts: 60 were stated in a cited report, 2 were stated only in part (a paraphrase, or a date taken from the report's own date), and 0 could not be found in the reports. Three more were the summary's notes about its own sources (for example that both reports are company press releases), not claims from them. No key fact contradicted a report. The reports of a twelfth event could not be fetched and it was left out.

Outlets, times and counts

An event's outlet count is the number of different sites among its reports, across every version of the event: a report's site is the address it was published at, so a reporter's byline and the outlet that ran it count once (behind an aggregator such as Google News, the outlet's own name is used). Sites that publish machine-generated stock scans (the MarketBeat network and alike) are not counted or shown, and an event built only from them is not on the site. Syndicated copies count: when a wire service or a group of stations runs one report on many sites, each site that ran it is an outlet (except the sites of one syndication network under one owner, such as the Big News Network's city sites, which count once together), so an outlet count measures how far an event travelled, not how many newsrooms reported it independently. Each event's own page therefore groups the copies and shows the number of distinct reports beside the outlet count. Its first report is the earliest publication time among those reports, in UTC, as the publisher's feed states it. These figures come straight from the reports' records, not from a model. Counts of events (per day, desk, month, category or name) count each event once, however many outlets carried it. When the pipeline opened two events for the same news (one deal written up twice, one launch over two days), the two are merged: candidate pairs on a desk within three days (by text similarity, a shared figure, or two or more shared names) are judged "same event or not" by a decision model (TypeSafe Jev). An event is merged only into an event it is itself judged the same as, with 85% confidence or more, within two days; merges are not chained, so a ruling and the reactions to it stay separate events (827 merges on October 11, 2026; each new event is checked against its published neighbours as it is published). Where write-ups of one event disagree on details, the model can judge them different, and a few duplicates remain. An event is filed on the day it entered the record, which can be a day or two after its first report. The events of the last three days are recounted every 15 minutes, as new reports arrive, and every figure on the site is recomputed from the whole record once a day (03:17 UTC). Reports judged not to be about the event are not counted. The outlet name shown is the site's own name when the name in the feed matches the report's address, otherwise the address itself: feeds carry bylines and aggregator names in that field.

Desks and categories

A classifier (from October 12, 2026, the decision model TypeSafe Jev, with DeepSeek V4.1 Flash for the reports it is less sure of; before, DeepSeek V4 Flash) files each report under one category, such as funding, regulation or earnings, and the pipeline routes it to one or more of the 17 desks. An event's home desk is the desk of its most-sourced write-up; the other desks list it with a link to that home. A second model reads each label against its event, and a label that does not fit is removed (414 of 12,084 on October 11, 2026, and every new event since). Categories are a model's judgement and can still be wrong; the trackers say which category or headline rule they read.

People, companies and bodies

The same model that writes the summary names the people, companies, products and bodies in each event. GetYourBrief maps each name to one of eight classes with a fixed rule, corrects places the model filed as companies (a list of countries, regions and U.S. states), and merges the same body under different names (for example "SEC" and "Securities and Exchange Commission") into one page. Reporters, commentators and private individuals named in the reports are named only inside the events: a person gets a page only as a public figure who is a subject of the news, judged per name by a decision model, and with the same bar as any name. The record names 35,290 of them; 1,918 have their own coverage page.

What gets its own page

Every event appears on the site. It lives as a section of its desk's day brief when that day carries four or more events, otherwise in the week brief. An event gets its own page when five or more sites ran it in three or more distinct reports, so that the page brings separate newsrooms together (693 events, including those that readers already reached from search, which keep their page). A name gets its own page when it appears in five or more events on three or more days (a person, only as a public figure; names readers reached from search keep their page, except people). One rule decides all of this, and the pages, the redirects and the sitemaps all read it.

Trackers

Each tracker counts the events of one kind across every desk. Layoffs are matched by headline (layoff, lays off, job cuts, cuts … jobs) and FDA approvals by headline. Funding rounds need both the classifier's category and a round named in the headline (a raise, a series, a seed, a valuation, a backer), because the category alone also caught loans, grants and market moves. Acquisitions and earnings likewise need the deal or the results named in the headline. The others use the classifier's category. A hand check of random rows on October 11, 2026 found the right kind of event in about 9 of 10 rows for funding rounds, layoffs, IPOs, acquisitions and earnings once the headline rules applied (acquisitions had been 7 of 10 and earnings 8 of 10 on the category alone); the misses are events the classifier filed under the category but that report something else (a market move, a forecast). Then a second model (TypeSafe Jev) reads every row and asks whether the event reports what the tracker counts: on October 11, 2026 it took 2,202 of 3,884 rows out (a forecast or survey about layoffs, a denied rumour, an appeal rather than a ruling, a summit filed as regulation, a coverage record) and merged 5 events counted twice; since then it reads each day's new rows. Every row links its reports, so a reader can check any of them. An event counts in one tracker only, headline rules first. A funding or layoffs row shows the figure its event's headline names (a dollar amount next to a word such as raises, secures, round or grant; a job count), only when the headline names exactly one, labelled as the headline's figure rather than a verified total. Treat tracker totals as a floor: an event the classifier filed elsewhere is not counted.

Who checks what, and how often

1. Summaries and key facts

   Checked automatically: Legal-safety screen on every sentence (how); structural and year checks; a second model review when they fail

   Reviewed by the data editor: A random sample against the cited reports; weekly, a second model re-checks a sample for legal safety (last: 2026-10-13, 200 sentences, 19 withdrawn)

   Last review: 2026-10-11

2. Day, week and month briefs, desk pages

   Checked automatically: Recent days recounted every 15 minutes, the whole record daily; counts compared with the record by the release tests

   Reviewed by the data editor: Three figures per page type checked against the reports

   Last review: 2026-10-11

3. Trackers

   Checked automatically: Membership rule applied to new events every 15 minutes, to the whole record daily; a second model reads each day's new rows

   Reviewed by the data editor: Rows sampled against the rule and the reports

   Last review: 2026-10-11

4. Name pages

   Checked automatically: Counts every 15 minutes; classes, merges, the public-figure rule and the page bar daily; a withdrawn name within 15 minutes

   Reviewed by the data editor: Class map and merges reviewed against a sample of names

   Last review: 2026-10-11

Reviews happen at each release of the site; the last review date is shown for each class and updated when a review is done.

The limits

GetYourBrief sees only the sources its pipeline polls, so an event that those sources missed is not here, and an outlet count is the count among them, not every outlet in the world. Publication times are the publishers' own and are sometimes late or wrong. Summaries can misstate a figure or a name; the reports are the record. Nothing here is financial, legal or medical advice.

How AI is used

Language models write the event summaries, key facts and names, and the category of each event; a decision model screens every sentence for legal safety, judges which write-ups report the same event and which people are public figures. GetYourBrief's publisher also uses AI tools to write the site's code and page templates. The numbers on the pages (outlets, times, counts) come from database queries over the reports' records; no model decides them.

Who answers for it

GetYourBrief's publisher is named on the about page. The data editor, a role at the publisher, answers for the method. To report an error, write to corrections@getyourbrief.com with the page address; the corrections page says what happens next.