Fact-Checking Policy
Checking happens before publication, not in response to a complaint afterwards. This page describes the sequence a claim goes through, so a reader who wants to test us knows where to push.
Our version of the developing-story problem is not the usual one. We are not racing anyone to be first. We arrive a day or two into a story that already has coverage behind it, so our danger is not haste but inheritance: picking up a number that was wrong the first time somebody printed it and has been copied ever since.
How a claim gets checked before it goes up
The writer’s draft carries its own sourcing. Every factual sentence has a note attached that the reader never sees, pointing at the document, the dataset, the recording or the interview it came from. An editor reads the draft against those notes rather than against the writer’s confidence, which is slower and the only way that catches anything.
A second editor repeats the check from the source material, not from the article, whenever the piece accuses a named person or company, rests on a document only the writer has seen, carries a casualty or damage figure, puts a number in the headline, or leans on a single anonymous source. That is a re-check, not a proofread, and it is the last stop before publication.
Alongside it runs the cold read, where an editor who has deliberately not looked at the source material marks every sentence they had to take on trust. It exists so a returning reader can follow the story, and it catches inaccuracy anyway: a sentence nobody can follow is often one the writer never pinned down.
The primary-source rule and the day-two trap
We go to the thing itself. The ruling rather than a summary of it, the filing rather than a note about it, the transcript rather than another outlet’s paraphrase, the person on the record rather than a description of what they think.
On a running story, five outlets carrying the same figure counts as one source until we establish whose figure it was. So we trace it: who published it first, what they attributed it to, and whether the original was an official statement, a spokesperson off the cuff, or an estimate that acquired a decimal point along the way. When it traces to one provisional statement, the article says so and warns the number will move.
Where no primary material exists, we attribute the claim to whoever makes it and leave it obviously secondhand rather than promoting it to fact.
We link sources in the sentence that makes the claim, so a reader can check us rather than trust us. Where a document cannot be published, the article describes what it is and how we came to see it.
Numbers, studies and the press release problem
A university press office is not the paper it is describing. Somebody here reads the study and checks how many people were in it, whether the headline finding was the question the work set out to answer or a subgroup noticed afterwards, and whether a link became a cause between the abstract and the release. Preprints are labelled as not yet peer-reviewed.
Business and motoring copy attracts softer numbers. Manufacturer range and economy figures are attributed to the manufacturer with the test cycle named, because they describe a procedure and not your driving. A percentage change is meaningless without its base, so we print the base, currency conversions get a date, and a company’s own adjusted measure of its performance is labelled as theirs.
Pictures and video
Nothing goes on the page whose origin we cannot establish. Images get a reverse search for earlier copies, a look at the file’s own data where it survives, and a check that the frame matches the claimed place and season, which means the signage, the language, the weather and the vehicles.
The standard failure is an older photograph recirculating as today’s, and it works because it is a real photograph of a real thing. Video gets the same treatment, plus the question of who shot it and how it reached us. A still whose provenance is no better than “supplied by” says so in the caption, and illustration made with software is captioned as illustration under our editorial standards.
Social posts, leaks and things readers send us
A screenshot is not evidence of anything. We go to the original post, check whether the account has a history or was created last month, and match the wording against an archived copy. If a post has been deleted we say we saw it and describe what we could confirm independently.
Some categories are claims by default. Transfer accounts in sport aggregate other people’s rumours, so a move gets reported when a club, a league or a named person says so. A datamined game file is evidence about what is in the file and no evidence about what a publisher will ship. An unattributed entertainment story about somebody’s private life that lands close to a release date gets the suspicion any other paid pitch would.
The same standard for everybody
A claim is checked the same way regardless of who makes it. A governing party and an opposition, a union and an employer, a regulator and the company it regulates: the question is what the evidence is and whether we have seen it, not whose interests the answer serves.
We will not set a checked fact beside an unchecked one for symmetry, and we do not use a line about both sides to avoid the work. Where a contested question is genuinely unsettled, the article says what is established and what is disputed and by whom, rather than splitting the difference.
When the checking fails anyway
It does fail sometimes. The fix goes on the article with a dated note saying what was wrong, and our corrections policy sets out who decides, how fast we reply and how the wrong version gets chased off our other channels. Errors that make it through are logged, and the log is read for patterns rather than filed.
We work to the IFCN principles on transparency of sources and method and on correcting openly. We are not a signatory and do not claim to be. The principles are a checklist we hold our own copy against, and if you think a specific article fails it, the address is support@thelittlebinger.com.
Last updated: August 2026.
