Four Ways a Number Goes Wrong, Taken From Our Own Corrections

Invention Is the Rarest Failure

If you edit anything that runs on numbers, you learn quickly that fabrication is not your main problem. Over the past month this publication corrected a run of figures before or shortly after publication, and almost none of them was simply made up. The rest were wrong in more interesting ways — and each way has a distinct signature you can learn to spot.

What follows is drawn entirely from our own work, including the pieces where the error was ours.

One: The Superseded Document

The most instructive error we made this month. A draft on the La Osa data center rezoning in Pinal County argued that supervisors had been asked to approve a 3,385-acre campus with no megawatt figure and no water volume anywhere in the application. It was a clean argument and the documents supported it.

They supported it for the original application. Three months before the vote, the project had been cut roughly 80% — from 59 data center buildings to 11 — and had published a 500 MW gas plant, a 1 GW demand cap and a 600 acre-foot water limit. The board that voted in August had numbers in front of it.

Nothing in the draft was fabricated. Every fact was sourced. The sourcing was to a version of reality that had been replaced.

The tell: a live application, permit, bill or filing that has been through revisions. Before characterising what a document contains, confirm which version the decision-maker acted on.

Two: The Figure That Drifted

A piece on the cyclospora outbreak cited 495 hospitalizations. The CDC’s own count, matched to the same date, was 356. The 495 came from a summary describing hospitalizations “overall” across a wider surveillance population — a real number, attached to the wrong denominator, that had been repeated until it looked like the outbreak figure.

Similarly, a naval logistics piece attributed 250-plus days at sea to two ships. It applied to one. The USS Abraham Lincoln had been deployed since November; the USS Tripoli did an ordinary six-month rotation and came home in July.

The tell: a number quoted without its date or its subject attached. “495 hospitalizations” and “356 hospitalizations as of 27 August, among outbreak-linked cases” are different kinds of statement, and only one of them can be checked.

Three: The Inverted Trail

We were asked to write up a story about a Detroit warehouse where the FBI found human remains in 2013 — real, prosecuted, thoroughly documented. The framing offered was that the trail led back to Tel Aviv and a well-connected family.

It led the other way. The convicted party was the American who ran the warehouse. The Israeli dental training centre was a customer, downstream of him, and nobody there was charged. The framing took a buyer and reinstalled it as the hidden origin, then supplied the family’s ethnicity as the explanation. That is an old template, and it was laid over a real case whose actual facts were freely available.

The accompanying figure — “over 1,200 human heads” — appeared in no court record or mainstream account either. Both traced to one blog.

The tell: a story where the causal arrow points at a group rather than at a defendant. Check who was actually charged. In fraud cases the answer is a matter of public record and takes about four minutes.

Four: The Number With No Parent

The strangest category. Widely circulated claims that humanoid robot production has passed 50,000 cumulative units originate with none of the companies involved. Tesla has never published an Optimus production count in any filing, delivery report or earnings call. The figure exists, propagates and gets cited, with no corporate source at any point in the chain.

The mirror image is a correction that never propagates. Amazon’s Blue Jay warehouse robot was retired in February; the disclosure is two sentences appended to the original announcement page. Dozens of articles still describe it as operational, because a launch travels and an update does not.

The tell: ask what document the number came from. Not which outlet — which document. If three sources cite each other and none cites a filing, you have one source wearing three hats.

The Habit Worth Building

These four have almost nothing in common except the outcome. A staleness check will not catch an inverted trail. Reading the primary source will not help if the primary source has been superseded. Tracing provenance will not tell you a figure has drifted from its denominator.

So the discipline is not “check the facts.” It is narrower and more mechanical: for every number that carries an argument, write down its date, its subject, and the document it came from. If you cannot fill in all three, you do not have the number — you have a sentence about it.

We publish with that gap sometimes, and say so when we do. The alternative is the failure mode this whole industry has settled into: fluent, confident, sourced-looking prose in which no individual claim can be traced to anything at all.