The NewsRupt

We would rather run five articles a week that people cite than thirty nobody finishes. Volume is not a goal and never becomes one. The weekly shape is deliberate: a Monday explainer, a Tuesday research or data piece, a Wednesday reference answer, a Thursday analytical feature or benchmark, and a Friday account of what actually changed that week.

Roughly thirty per cent of a month's output is explanatory, twenty per cent news analysis, twenty per cent original research and benchmarks, fifteen per cent maintained reference answers, ten per cent market maps, and the remainder interviews. We monitor that mix on a rolling thirty-day basis. When it drifts toward commentary on other people's announcements, we stop and commission measurement instead.

At least ninety per cent of what we publish must stand on its own reporting, measurement or analysis rather than on reacting to a press release. Where a piece involves a company affiliated with this publication's owners, the affiliation is disclosed in the piece itself, at the point a reader needs it, not in a footer. No placement in any ranking has ever been sold, and none is for sale.

AI systems draft, summarise, extract and structure. They do not decide what is true. Every factual claim that survives to publication has been checked against a primary source by a person, and every quotation traced to its origin. Automated wire ingestion files stories as drafts for review; it does not publish.

The production chain is fixed: idea, research, AI draft, source verification, human edit, fact check, approval, publication, scheduled review. A piece that cannot clear source verification is killed rather than softened.

Primary sources first: filings, specifications, system cards, datasets, court documents, the announcement itself. Where reporting is the source, we look for two independent outlets and avoid circular sourcing — three publications citing the same unnamed person is one source, not three. Where reporting is contested or a bill has not been filed, we say so in the sentence rather than rounding it up into a clean claim.

Fact and analysis are visually and grammatically separated. We do not manufacture consensus, and we do not balance a documented fact against an unevidenced objection for symmetry.

Our research franchises are recurring, which is the point: the same instrument run again produces a comparable number. Every edition publishes its method, its frozen prompt set or instrument, its limitations, and the raw data as CSV and JSON under CC BY 4.0 with a suggested citation. If our method is wrong, the data we ship is what proves it.

Every edition states what it cannot claim. A measurement without stated limits is marketing.

When a new development overlaps an existing page by more than about seventy per cent, we update that page rather than publish a near duplicate. Reference entries in Answers carry three dates: first published, last updated, last reviewed. Pages older than ninety days enter a review queue, as do pages with broken links, superseded model names or stale pricing.

Ranked listicles as a default format. Headlines written for a keyword rather than a reader. Pages that exist because a phrase has search volume. Undisclosed affiliate arrangements. Republishing a vendor benchmark as though we had run it. Any claim that a number is definitive when it moved the last three times we measured it.

We want to be retrievable and quotable by search engines and AI answer systems, and we pursue that only in ways that also make the page better for a person: a clear answer near the top, honest headings, structured data that matches the visible content, linked sources. Optimisation never decides what we publish or how a finding is characterised.