Fact-Checking Policy

Checking a story before it publishes is a different discipline from correcting one afterwards, and it is the part of the job no reader sees. This page describes what it consists of, including where it gets difficult.

Last updated: August 2026.

Who checks what, and in what order

The writer checks first, and the standard is that every factual sentence can be traced to something they have in front of them. Figures, dates, job titles, legal outcomes, the spelling of every surname. Reporters keep the source for each claim beside the draft, which adds about ten minutes to a normal news story.

Then an editor reads the draft hunting for one thing: a sentence with nothing behind it except a single source. What survives is either supported twice or flagged in the copy as unconfirmed. A second editor reads before publication when the piece accuses a named person or company, when it rests on a document only we hold, when numbers drive the story, or when the legal risk is obvious.

Documents before descriptions of documents

We would rather have the filing than a summary of it, the regulator’s database entry than the manufacturer’s press release about it. A press release is evidence of what an organisation wants said, which is not the same thing as the event it describes.

On-record interviews carry weight in proportion to how close the person was to what they describe, so somebody who read the document beats somebody who was told about it. Where we hold a document we cannot publish, we say what it is and how we satisfied ourselves it is genuine. Attribution and anonymity are set out in our editorial standards.

Links in the copy, not a bibliography at the bottom

A claim resting on a public source links to it in the sentence where it appears, so a reader can check us without writing in. The link goes to the filing, the register or the dataset itself, not to another article about it.

Where a source cannot be linked, because it is behind a paywall, unpublished, or a person we agreed not to name, the copy says which of those applies. We do not cover the absence of a link with a vague verb.

Pictures, clips and screenshots from strangers

Nothing runs as evidence of an event because it looked convincing in a feed. An image gets a reverse search in case it has been circulating for six years under a different caption, a look at whatever metadata survived being uploaded, and a check of the details that place it: the language on the signage, the weather, the shadows.

Where a picture or clip came from a social account, we work out who posted it and whether the account existed last month, whether the clip is the earliest copy or a crop of something older, and what the person claims to have seen rather than what the caption asserts. We ask whoever took it for the original file and for permission, and the caption credits them. Handout images from a company are labelled as handouts, and we publish no synthetic image as a photograph.

Screenshots are the easiest thing in the world to fake, so a screenshot of an email, a patch note or a deleted post is checked against something else: the archived page, the sender’s own copy, a second recipient. If it cannot be corroborated it does not run, even when it is the best item on the list.

Numbers, studies and the press release about the study

Statistics get taken back to the table they came from: the original dataset rather than somebody’s chart of it, the base the percentage was calculated on, and which comparison year was chosen. A number that has passed through three retellings has usually gained a decimal point or lost a qualifier.

With research we read the paper instead of the summary. Sample size, who paid for it, whether it was peer reviewed, whether an animal study is being written up as a human finding, and whether the authors themselves claim what the headline claims. In Lifestyle and Science, the gap between a real study and the supplement marketing quoting it is where most bad health reporting lives.

The same test regardless of who is talking

A claim is checked to the same standard whoever makes it: a minister, a party whose analysis we happen to share, a company we have criticised before. We do not soften the check for a sympathetic source. That is the failure mode that catches sceptical publications, because a sceptic’s own assumptions never feel like assumptions from the inside.

Sceptical does not mean symmetrical either. Where the evidence sits heavily on one side we say so rather than build a balance out of a rebuttal nobody can support. Two positions get equal room when both are supported.

The developing story, where all of this gets hard

Speed is where checking fails. A story is moving, other outlets have gone with it, and the honest position is that we know three things and can confirm one. So we publish the one, name where it came from, and write plainly that the rest is unconfirmed. Casualty figures, arrest numbers and the cause of an outage wait for the body that would actually know.

An editor sets a floor on each running story, the minimum that has to stand up before we post. Nothing publishes because a competitor has it. Being forty minutes late has never cost us a reader worth having, and going first on a figure that gets revised twice costs us the only thing we own.

When the check fails anyway

Some of this will not work. A source will turn out to have been confident and wrong, and a document will be genuine and misleading at once. When a published claim does not hold, the article is corrected on the page with a dated note. Who decides that, and how quickly, is in our corrections policy.

We also go back through the checking to find where it broke, because one wrong story is usually a process that let it through rather than a careless afternoon. We work to the principles of the International Fact-Checking Network on sourcing and corrections, without claiming a badge we do not hold. If a check described here was not done on something we published, write to support@thekissingpunk.com and say which one.