Our approach
A piece can match your interests and still have little to say. Brevva treats the quality of the work and its personal relevance to you as separate questions.
We look for work with something to say and support for its claims. Your interests, follows and feedback help decide which pieces to choose for you. A score helps with selection; it does not settle whether a claim is true or whether a piece is right for you.
We ask nine questions about the work itself. An AI model gives a reason and a grade for each answer, then a fixed rule combines the grades. Clear writing matters, but polish should not make a thin argument look substantial.
Substance asks whether a piece advances understanding. Originality looks for reporting, analysis or synthesis that adds a fresh angle. Primary contribution rewards firsthand work, while strong analysis of others’ reporting can still be valuable.
Factual support asks whether central claims are specific, checkable and supported by the available text. Proportionate framing asks whether the headline and presentation fit what the piece actually says.
Attribution asks whether sources and claims can be traced. Author accountability asks who stands behind the work. Disclosure of interests asks whether relevant stakes are made clear.
Clarity asks whether the work is readable and well structured. It carries less weight than substance; fluent writing cannot rescue an empty argument.
Four separate checks can lower the combined result: a demonstrably fabricated central claim, undisclosed promotion presented as editorial work, a materially deceptive headline, or no attribution where factual claims require it. These are model-assisted judgments, not a guarantee that every false claim will be caught.
The documented rubric v2.1 and f-default-v1 rule give each graded part a value from 0 to 1. The combined score weights credibility at 65%, transparency at 25% and craft at 10%.
Within credibility, substance carries 30%, originality 25%, factual support 25%, primary contribution 10% and proportionate framing 10%. Within transparency, attribution carries 45%, author accountability 30% and disclosure of interests 25%. Clarity supplies all of the craft grade.
When a grade is missing, the remaining grades in that layer are reweighted to total 100%; a layer with no grades contributes zero. If any of the four separate checks above is triggered, the rule sets the combined score to 0.10. This describes the documented combination rule; it does not establish which version assessed every item you see. Discovery and your follows and interests also affect what appears in a briefing.
Quality is one question; whether a piece belongs in your briefing is another. Brevva answers the second in stages.
Sources. Brevva follows newspapers, magazines, independent writers, newsletters and podcasts. A newly found source is assessed in stages, first for relevance to readers, then on a sample of its work, before anything it publishes can be selected. A source you add yourself skips that assessment.
Reading. A piece that passes a first screen on its headline and summary is analysed on its text; a very long piece is read from its opening. With too little to assess, a piece is held. Some subscription sources offer only an abstract or excerpt: those pieces are analysed on what is visible and count for less in scoring, and any note says what it was written from. A piece can also appear without a note, when there was too little to write from or a note could not be produced.
Six dimensions. Each analysed piece is scored for you on six separate dimensions: zeitgeist (what is moving now across the sources), familiar (your established interests), similar (close to what you have valued before), surprising (what you would not have gone looking for), urgent (what will not wait) and challenging (what pushes back on your view). Scores are compared within that day's pool rather than against a fixed scale, so a quiet day and a busy day each produce a briefing.
The second look. The strongest candidates then go through a separate pass that asks a plainer question: would this reader find this interesting? A high score on topic is not the same as being worth your time.
Composition. The edition is assembled in sections: top stories, your topics, the writers and publications you always read, podcasts and video, and something unexpected. One source cannot fill the briefing, near-duplicates are removed, pieces you were recently sent are not sent again, and no single theme is allowed to crowd out the rest.
Your feedback. Ratings on a topic, a source or a writer, your follows and your plain-language preferences change the scores that drive the next selection; a writer or publication you choose to always read is given its own place in the edition. The balance between the dimensions differs per reader and is tuned over time; this page describes the mechanism, not a guarantee about any single pick.
Checking whether claims appear supported by the material available to us is not independent fact checking against the world. Sometimes we can see only an excerpt, description or publisher summary. A briefing should say so rather than imply that we read the whole work. A quality result also cannot establish that an idea was original before every other publication, guarantee access to a paid source, or promise coverage of every topic.
Your interests and follows shape which available pieces belong in your briefing. A followed writer is your deliberate preference; it does not mean every piece from that writer passed the same discovery rule. Discovery, source availability, duplication and a mix of formats affect the final selection. When we show a comparison count, it covers only the selection pool recorded for that period, not everything published or what you actually read.
In a July 2026 study, we compared published selections from three curators with contrasting articles drawn from an eligible source pool. The rubric ranked selections from The Browser and Arts & Letters Daily above their contrasts more often than chance. It did not establish that separation for Marginal Revolution. Curators may select for surprise or interest that this quality rubric does not measure well.
These were curator selections, not independent human ratings of every item. The comparisons were modest and could also reflect genre or article length. Overlap between the curators was too small to test whether they agreed on quality.
A separate July 2026 set of small, controlled writing changes tested reactions to padding, polish, confident wording, apparent authority and AI-style prose. It caught an inserted fabricated claim and showed instability in some individual judgments. Some transforms changed more than style, and later exploratory checks used different conditions. Those results guide revisions; they do not certify every future article or prove that model bias is absent.
Each briefing links to its sources when an appropriate public link exists. Readers can rate items to help shape future choices. If something looks wrong, reply to a Brevva email and tell us. Reporting it does not change the original source. Our general sample is the same for every visitor; an individual briefing follows that reader’s interests.
Request an invite and we’ll email you when yours is ready.