A fact-checking desk does not decide whether a statement sounds plausible. It works backwards to whatever the claim rests on and establishes what that source can and cannot support.
The claim has to be made checkable
Only statements about verifiable facts can be checked, which excludes predictions, opinions and expressions of value.
A vague assertion is first narrowed into a specific proposition, since a claim that could mean several things cannot be assessed as any of them.
Checkers publish the version they assessed for exactly this reason, so a reader can see which interpretation was tested.
Primary sources outrank everything
The first step is finding where a number or quotation originated, tracing back through the chain of articles that repeated it.
That chain frequently ends somewhere weaker than expected: a press release, a preliminary document, or an aside in an interview.
Numbers also change meaning as they travel, and a figure describing one narrow group is regularly reported as though it described everyone.
Context is checked alongside accuracy
A statement can be literally true and still create a false impression, through selective comparison or an unusual baseline year.
Checkers therefore assess whether the framing supports the conclusion drawn, not only whether the individual figures are correct.
This is why ratings scales include intermediate positions, since a binary judgement would misdescribe most of the claims that circulate.
Images and video need different tools
Visual claims are usually not fabricated but recycled, showing a real event presented as a different one.
Reverse image searching, examining shadows, signage and weather, and comparing a location against mapping imagery all establish where and when something was recorded.
Synthetic media has made this harder, and the practical response has been to prioritise establishing where a file first appeared over analysing the file itself for signs of manipulation.
The verdict is published with its workings
A check sets out the sources consulted, the responses of anyone whose claim is assessed, and the reasoning that produced the conclusion.
Where evidence is insufficient, the honest outcome is that a claim is unsupported rather than false, and conflating the two is a recognised failure.
Corrections are handled the same way, with the original text retained and the change recorded, so that a check which turns out to be wrong leaves a visible trail.
Publishing the workings is what allows the check itself to be checked, and it is the only thing separating a fact-check from another assertion competing for belief.