How to Check a Claim

The order matters more than any single step

Every page in this section is a check of some kind. None of them is the method, because the method is the order you do them in, and the order is what decides whether you are testing a claim or performing a test on it.

What follows is eight steps. Each one has a page behind it, so this is the shortest thing in the section and the one that points at the most. The last step is the one that decides most arguments, and it is almost never stated.

1. Say what would change your mind, before you look

This is the only step that has to come first, and the only one that costs nothing.

Write down, before any searching, what result would make you drop the claim. If nothing would, the exercise that follows is not a check. It is a search for confirmation with a check's vocabulary attached, and it will succeed, because searches of that kind always do.

The reason it always succeeds is arithmetical and it has a page of its own. A claim of the form these numbers match, and that is unlikely has two halves, and only the second can be examined. The first — that the numbers match — is usually true and usually uninformative. What makes it uninformative is the figure such claims never supply: how many chances the search had.

2. Count the chances

If a method can be applied to a name, a date, a building and a sum of money, and any of four operations can be performed on each, and a match on any of them counts, then the search had dozens of opportunities and not one. The probability that something matched approaches certainty long before the claimant has done anything wrong.

The control case is the one to keep in mind: the same technique that finds assassinations encoded in a religious text finds them in Moby-Dick, which nobody claims is encoded. A method that produces hits in a control is not producing evidence when it produces hits elsewhere.

3. Get the document, not the article about the document

Almost every claim in this archive rests at some remove on a document, and the remove is where the damage happens — a figure becomes approximate, a date slips a year, a conclusion acquires a certainty the original did not carry.

If the document is governmental and American, you can usually ask for it yourself, and the procedure is its own page. Two provisions do most of the work: since 2016 an agency must identify a foreseeable harm to withhold under a discretionary exemption, and the deliberative-process exemption cannot be applied to records more than twenty-five years old at all. Both are free wins on an appeal and neither is widely used.

4. Read the markings before you read the text

When the document arrives, the body of it is the second thing to look at. The markings tell you more, and they tell you something the prose cannot: how the people who wrote it rated their own material.

The banner at the top is the least informative part, because it reflects the highest classification anywhere in the document. The portion marks beside each paragraph are the useful ones. If the paragraph containing the damaging claim is marked (U), it was never classified, which means the agency did not consider that sentence a secret — and a document's own markings are a harder thing to argue with than anybody's summary of it.

5. Date the claim, not the retelling

A prediction is only interesting if it was made before the event, and a surprising number were not. The thing to find is the earliest dated appearance in a source you can point at, which is frequently later than the claim's own account of itself.

The register of dated predictions exists partly to make this habit visible: it carries a lead-time column, measured from the claim entering circulation to the date it named, and the spread runs from four centuries to the same year. The ones made the same year are not less interesting. They are a different kind of claim, and being able to tell them apart is the point.

6. Ask what the organisation publishes about itself

For claims about a body — that it controls something, that a named person belongs to it — the fastest discriminating question is not whether the claim is plausible. It is what the body publishes, because that determines whether any answer is available to you at all.

Fourteen of them are set out here with that column filled in, and the distribution is not what the subject leads you to expect: of the eleven with a membership that could be disclosed, ten are knowable by some route and exactly one publishes nothing at all. Several publish a current roster with names. One publishes its roll a century late. Two had theirs published because a government seized the papers.

7. Ask who owns the sources that agree

Independent corroboration is the strongest thing a claim can have, and the hardest to assess, because agreement between outlets is only evidence when the outlets are actually separate.

That is an ownership question with a factual answer, and the chart is here. It is also the step most often skipped in both directions: by people who treat five reports as five confirmations when they share a wire service or a proprietor, and by people who treat common ownership as proof of coordination, which is a different claim requiring its own evidence.

8. Separate the measured effect from the inference built on it

The hardest case is not the claim with nothing behind it. It is the claim with something small behind it, carrying something large on top.

The ledger of panics has three rows of exactly that shape. A 1941 pediatric study of two hundred children really did find more than ordinary nervousness in 76 per cent of the heavy radio listeners against 40 per cent of the children who listened to nothing — and could not separate the listening from the crowded flats and absent alternatives the same paper describes, and made no finding at all on the delinquency that was the actual charge. The American Psychological Association's 2020 resolution says both halves in one breath: insufficient evidence of a causal link to violent behaviour, and a small, reliable association with aggressive outcomes of the yelling-and-pushing kind that is difficult to extend to violence.

The question to ask is never is there anything to it. It is how big is the thing there, and does the claim fit on top of it.

And when the answer is that a programme was alleged for years and then documented, that is worth knowing too, because it happens: seventeen of them are on one timeline, with a median gap of twenty-eight years between operation and public record.

The step that decides most arguments

Here is the part that is almost never said. Most of the checks above can confirm a claim and cannot refute it.

Take the commonest claim in this subject: that a named living person belongs to a particular organisation. If they do and they have said so, you will find it in minutes. If they do and have not said so, you will find nothing. If they do not, you will also find nothing. Two of those three states produce an identical result, so a search that comes back empty has told you almost nothing — and the claimant can say so, accurately, for as long as the argument lasts.

That asymmetry is structural, not rhetorical. The Freemasonry file is the clean illustration: the English registers from 1751 to 1921 are online and searchable, about 1.7 million names, published by the craft's own museum — and they stop a century short of anybody living. For current members the grand lodges publish constitutions, officers and addresses, and whether an individual is named is left to him. Only two mechanisms have ever produced a list of living members, and neither belongs to the organisation: a state seizing the papers, as Italy did in 1981, or an employer compelling a declaration — which the Metropolitan Police now does, and which the High Court declined to disturb on 17 February 2026.

So before running a check, ask what a negative result would mean. Sometimes it means a great deal. Whether a given machine on the internet is really Google's crawler, for instance, is fully decidable: Google publishes the address ranges, the name has to resolve backwards and forwards to the same place, and an address that fails both is definitively not what it claims. Checks like that are rare and worth recognising, because they are the only ones where we looked and found nothing is an answer rather than a shrug.

What this page is not

It is not a way of settling anything by itself. Every step points at a page that does one job, and the pages are where the work is.

It also does not tell you what to believe when the method runs out, which it will. Several entries in this archive end with a question the available record cannot close — and the honest response is to say which part is established, which part is not, and what evidence would move it. That is a worse feeling than a verdict and a more accurate description of the situation.

The thing to take from it is smaller and more useful than a verdict: write down what would change your mind first, count the chances the search had, and know before you start whether finding nothing would mean anything at all.