Voice & quality

Quality gates and fact-checking

How Mastheads checks every article - sources read before writing, every claim tested against them, eleven recorded checks - and how the article shows you all of it, including what fell short.

Every article shows its work. Open any article and you can see what it was written from, which checks it passed, which it failed, and what the desks could not fix. This page explains each of those checks, where they happen, and what they mean for what you publish.

The checks have been refined across more than 20,000 published articles. They do not replace your judgement; they show you exactly where to use it.

On this page

Before a word is written

The sources are read in full. The Source Desk opens every page it chooses and reads all of it, not a snippet. Copies of the same report on different sites count as one source.
Outlets you have ruled out are never used. A domain's ignored outlets are never fetched and never cited.
A topic with no usable sources is not written. If the web has nothing usable, or what it found is about something else, the run stops, nothing is charged, and the article says why.
You set a minimum number of sources. The default is two. The desks aim for it, and the article is delivered either way - but a shortfall is shown in amber, for example "Only 1 source of the 2 you asked for", never as a pass.

While the article is checked

Three desks read the draft, and each fixes what it finds in one pass:

The Fact Desk checks every number, date, name and quotation against the pages the desks read. With Research on, a fact must trace to one of those pages. One that does not is cut, made general, or marked for your review. Nothing is invented to fill the gap.
The QA Desk checks the article delivers what its headline promises, and that its search title and description describe the article honestly.
The Editor-in-Chief reads the whole piece fresh and rewrites unclear or crowded paragraphs, without moving a fact, a name or a link.

Then the Standards Desk reads the finished article against the house standards every article meets, and repairs what it can.

Links you can trust

Every link in an article goes to a page the desks actually fetched and read. Any other link is turned back into plain text before you see the article - so the article cannot cite a page that does not exist.

Eleven recorded checks

Every article carries eleven measured checks, shown on its authenticity panel:

Check	What it measures
Readability	How hard the article is to read, against the reading level you set
Source overlap	How much of the article matches one of its sources word for word
Banned phrases	Stock phrases from the house list
Paragraph shape	The length and shape of each paragraph
Talking about itself	The writer describing the article instead of writing it
Clean ending	That the article ends properly, not mid-thought
Contractions	Contractions used where they read wrongly
Currency style	Money written the way the country writes it
Quote marks	Quotation marks correct for the language
Country spelling	Spelling correct for the country - colour or color
Tone of voice	That the article matches the tone you set

Readability is graded with the Flesch-Kincaid scale in English and with LIX in other languages.

Each check has three possible answers:

a tick - it passed
amber - it failed, and the panel says how
not recorded - it could not run, for example on a very short article or with no sources to compare against

A check that could not run never shows a tick.

What we could not fix

When a desk finds a problem it cannot repair, it writes it down. The panel lists every one of them under What we could not fix, in the desk's own words, with what it would take to fix and which desk raised it. The panel header counts them, so you know at a glance whether there is anything to read.

Desks that did not work are shown as such

The panel lists the desks that worked on the article. A desk that stood down - for example the Fact Desk on an article written with Research off - or one that failed at its own job, appears without a tick. A run is never shown as more checked than it was.

Needs a look

The Standards Desk holds an article as Needs a look only for real risk:

a person or organisation named that no source mentions
a policy concern raised by the Fact Desk
an article that stops mid-sentence or wanders off its own subject
an order the sources could not fill - for example fewer entries than the headline promises
our own working notes left in the text
a long quotation that matches nothing in any source, or an article where none of the figures appear in any source
an article in the wrong language
a section you set to always cite that came back with no link

Smaller issues are fixed or recorded, not used to hold the article. A held article is never published automatically. You choose Regenerate or Publish anyway. See Review and approval.

No two articles the same

The shape of an article - its opening, its close, its structure - varies from one article to the next, so a run of fifty does not read like one template filled in fifty times. The headings you pin yourself are always kept.

What the checks do not promise

No set of checks makes an article correct on its own. A figure can appear on a real page and still be attached to the wrong product or year. That is why every claim links to its page, and why the panel exists: it shows you what was checked and where to look, so the final call - and your name on the article - stays yours.