Methods & quality

• Source. The published abstract and title, nothing else.
• Composition. AI models read the abstract and fill a fixed structure: study type, population, groups, the main result, further results, safety, the authors' conclusion.
• Checks. Every number must appear in the abstract, the authors' conclusion must be verbatim, a ratio shown with both groups' figures must run the way those figures run, and a second AI reading checks that each value stands with the group and the outcome the abstract gives it.
• Left out, never guessed. A line the checks cannot confirm is left out, and the page says so. A main result that cannot be confirmed holds the whole Brief back.
• The editor sets the rules, decides the questions the checks cannot, audits published Briefs, selects the Featured studies, and corrects errors readers report.

The full text is not used: most of it is behind subscriptions, and a summary of what every reader can check is worth more than one of what they cannot. The abstract's limits are the Brief's limits, and every Brief links to the original article. An article without an abstract gets no Brief. Society statements and guidelines are summarised without a results section, as their own type.

• Study type and design, as the abstract describes it.
• Population and groups, labelled as the abstract labels them.
• The main result: the end point the abstract itself calls primary or main. When the abstract does not report its result, the Brief says so where the Brief's text allows, and never fills it from another result.
• Further results and safety, each number as printed.
• The authors' conclusion, in their words.

Nothing is rounded, converted or recalculated. In Briefs made since 8 October 2026, an abbreviation is written out at first use only where the abstract itself expands it; older Briefs in the archive may carry an expansion from a standard list.

Every new Brief passes the checks in force on the day it is made. The checks have been tightened several times, most recently in October 2026, and a Brief is not re-checked when a rule changes after it: Briefs in the archive made before 8 October 2026 passed an earlier version.

The checks on numbers, the conclusion and direction compare the Brief's text with the abstract's by fixed rules, so the same Brief always gets the same answer. The Brief itself is written by an AI model and can differ from one reading to another; what cannot differ is that every number in it was printed in the abstract. A second, separate AI reading then checks that each value is attached to the right group and outcome and that no main result is missing.

What that second reading finds is settled in one of three ways. A question about which group a value belongs to, which way a result runs, which two groups are compared, or whether a value is right at all goes to the editor, and the Brief waits until it is decided. Any other objection is settled by the reading itself: it may keep the line only with a verbatim quote from the abstract, or replace it with a correction whose numbers must again be printed in the abstract; otherwise the line is left out. A result the abstract has but the Brief missed is never added afterwards. A Brief whose study type cannot be confirmed is not published.

A line that fails a check is first composed once more from the abstract, with the failure named. If it still fails, it is left out, and the page carries a note that parts of the Brief are withheld, so a reader knows something stood there. A main result, a safety line or a statement that still fails is not left out: the whole Brief is held for the editor instead. A value is drawn as a bar or a plot only when it also stands as text beside it.

The editor writes the rules that decide which articles get a Brief (see What's included) and the rules the checks apply. Where the abstract leaves a question the checks cannot answer — which group a value belongs to, which way a result runs, which groups are compared, whether a value is right — the editor decides it.

CurrentCardio Weekly and its Featured studies are the editor's selection, made from a shortlist an AI reading proposes without seeing any study's results. The shortlist is judged on the question, the design, the population and the end point, and also on the size of the study and the journal it appeared in; a study's result and its sponsor play no part. When the editor has not published an issue by its deadline, it goes out with the editor's choices topped up from that shortlist. My Week changes which Briefs a reader sees, never what a Brief says.

The method is measured against a reference set of Briefs checked by hand. No text from that set is shown to the models, but what it revealed shaped the rules: each round of measurement ended in the editor's decisions, and those decisions became general rules. Before the method was used on the archive, whole quarters of real articles from four journals were audited against their abstracts, Brief by Brief, and the editor decided on the findings. Confirmed errors are kept as cases the method is replayed against; not every case runs automatically on every change.

None of this makes a Brief infallible. Always read the original article before relying on a finding.

The journals are checked every hour. A Brief usually appears within a day of the abstract; one with a question for the editor appears once it is decided. An abstract that is published more than a week after the article is first listed can be missed.

Found an error? Write to contact@currentcardio.com with the Brief's address.