Fact-Checking Policy
This page describes how a claim gets checked before it reaches a bulletin, and what we do when the check was not enough. It names steps rather than intentions. Last updated: August 2026.
Who checks what, before it publishes
The writer checks their own copy first, line by line, against the source. Every figure, name, title, date and quote has to trace back to a document or a person, and the sources stay in the file with the draft so an editor can open them rather than take the writer’s word for it.
Then two people read it. The editor on the shift checks the copy against those sources. A second editor who has seen none of the source material reads the bulletin cold, looks away, and says out loud what happened. If they cannot, it goes back. A second, independent check of the documents is required when a bulletin alleges wrongdoing by a named person or company, reports a death or casualty figure, rests on an anonymous source, or carries a number off a story still moving.
Primary sources first
The filing, the docket entry, the transcript, the regulator’s own release, the club statement, the dataset. Those are sources. Another outlet’s report is a lead, and so is a screenshot, a post quoting a document, or an aggregator’s summary of one. The job is to find the original, which on most days is public and takes ten minutes.
Where the primary document exists and we cannot get it, the bulletin says what we relied on instead and who reported it first. That is a weaker bulletin and it reads like one. On-record interviews outrank background conversations, and written confirmation outranks the same thing said by phone, because we can quote the first accurately a week later.
Links, so you can check us
Claims carry their source as a link inside the sentence making the claim. We link the document rather than a story about it, and we link to the page or section where the line actually sits where the format allows. Where a source is a PDF or sits behind a paywall we say so before you click.
Some sources cannot be linked: a document not published online, a dataset we were sent, an interview on the record. In each case the bulletin describes the source precisely enough for you to request it yourself. Where another outlet did the reporting we credit them and link out, aware that the click leaves us. Anyone wanting the material behind a bulletin can ask at support@theibulletin.com.
Numbers, studies and surveys
Numbers are where a short bulletin fails, so they get the most work. We open the original data rather than the summary written about it, then check the base and the period, because a rise against a quarter and a rise against a year are different facts wearing the same word. Provisional figures are called provisional. We keep percentage change and percentage points apart, and we never round in the direction that improves the story.
For a study we check who paid for it, whether it is peer-reviewed or a preprint, how many people are in it, and whether the finding sits in the whole sample or in a subgroup found afterwards. Then the question that kills most of them: does the claim in the release follow from the result? A survey commissioned by the company it flatters does not run here.
Pictures and video
No image runs unless we know it shows the event we are attaching it to. That means a reverse image search to find the earliest appearance, a look at whatever metadata survives, and a check of ordinary physical detail: light, weather, signage, plate formats, the language on a shopfront. Footage rotated, mirrored, cropped or reposted through three accounts stays unverified until the original is found.
Handout images are labelled as handouts and the supplier named. File pictures say when they were taken. We contact whoever filmed something before using it, ask for the original file rather than the compressed copy, and ask whether they shot it themselves. We publish no generated imagery of real people or events, a rule set out with our other tool rules in our editorial standards.
Social posts and user-generated material
A post is quotable when we know who is behind the account and that person is in a position to know what they say. Follower counts, verification badges and how fast something spreads count for nothing. Where an account is anonymous and the claim cannot be corroborated elsewhere, it stays out of the bulletin however many people have already seen it.
Screenshots need archive evidence before we treat a deleted post as having existed. Official accounts get checked against the organisation’s own site, because a convincing impostor on a fast story is routine rather than rare. On a developing story the most-shared clip is reliably the least verified, and thirty minutes in the safe assumption is that the video going around is from somewhere else.
The same test whoever is speaking
A claim is checked the same way regardless of who makes it, which party they belong to or which government they speak for. We do not print a false claim in a headline and correct it four paragraphs down, and we do not set a verified fact beside an unverified one and call the pair balance.
The practical version is narrower than the principle. The claim needing the hardest check is the one that would make the better bulletin, because that is the one waved through. So we name it on the desk, and the person who wants the story is not the person who clears it. We work to the IFCN principles on accuracy, sourcing and openness about method, without claiming a membership we do not hold.
When the check fails anyway
Sometimes it does. A document read too fast, a body confused with another, a figure that was right at nine and revised at ten. The fix happens on the bulletin itself with a dated note, on the terms in our corrections policy, whether the error started with us or with a source we trusted.
We also record where the check broke, not only that it did, because a missing step is a different problem from a step taken badly. If you can see the join in something we published, quote the sentence and write to us. Readers find our errors faster than we do.
