Welcome to the Receipts Desk, where the numbers everyone is sharing get checked against the documents they claim to come from. Twelve days before a midterm election, this desk expected its first entry to be an argument about arithmetic. Instead it is an obituary, for a polling firm that never existed and the numbers it never collected.

A firm calling itself Wedgewood Polls appeared this year, seemingly from nowhere, and began releasing surveys on the Senate races that will decide control of the chamber: Georgia, Michigan, Texas, and more. Beginning in June, its numbers entered the bloodstream of American political coverage. The New York Times included Wedgewood in its polling coverage. USA Today wrote its results into stories. Nate Silver’s Silver Bulletin, the successor to the most famous polling average in the country, folded Wedgewood’s figures into its averages, where they quietly nudged the numbers that campaigns, donors, and journalists treat as the state of the race.

None of it was real. On October 20, Wedgewood admitted in a public letter that its results were entirely fabricated. The firm described the deception as a social experiment, and stated its hypothesis in its own words: “How easy is it for an entirely fabricated firm to gain traction, be taken seriously, and expose how little scrutinization exists in media and public commentary?” The experiment is over, and the hypothesis held. The answer was: effortless, and none.

Credit where it is due: the fraud was not caught by any of the institutions that amplified it. It was caught by FiftyPlusOne, a small polling and elections website that got a tip, paid to recreate one of Wedgewood’s recent polls, and found that the answers came back fundamentally different and that the way respondents were identified did not match. Recreating the work is how you check a number. It is also, apparently, a step too laborious for the largest newsrooms in America, which preferred to ingest the data and move on.

The cleanup tells you how deep the contamination went. The New York Times has removed 15 Wedgewood polls published between June and October. USA Today is attaching editor’s notes to its stories and removing references. The averages that included Wedgewood cannot simply un-ring themselves, either: for months, every average that carried its numbers was, to some degree, a work of fiction, and nobody can reconstruct what those averages would have said without it.

And the numbers did work while they lived. At least two Democratic candidates campaigned off the fake data. Mississippi Senate nominee Scott Colom promoted a Newsweek article, published October 14, built on a Wedgewood poll showing him within striking distance of Senator Cindy Hyde-Smith. Newsweek has since retracted it. Colom’s campaign says it had no role in commissioning or conducting the poll and only referenced it as a public survey, which is precisely the problem: a fabricated number, laundered through a famous magazine, becomes a fundraising pitch and a morale story before anyone asks who collected it. Jim Kessler, a former staffer to Chuck Schumer, said the quiet part plainly: Wedgewood “committed some sort of fraud,” and “it was designed to affect the outcome.”

Think about what the polling industrial complex actually sells. It does not sell surveys. It sells certainty, packaged as an average, updated daily, precise to the decimal point. The average is only as honest as its weakest ingredient, and the system has no immune response: a new firm with a name, a website, and a spreadsheet was admitted to the average on sight, because volume is rewarded and auditing is expensive. Wedgewood’s founders say they wanted to expose that. They have. The firm’s website and email are now dead, its proprietors unreachable, and the exposure is complete enough that it should embarrass everyone who touched the data.

So, twelve days out, here is the Receipts Desk rule for every poll you will see between now and November 3. Before you share the number, ask who collected it, when the firm was founded, and whether anyone has ever recreated its work. If the answers are a name you learned this summer, a post office box, and silence, you are not looking at a measurement of the electorate. You are looking at content. The electorate gets measured on Election Day, by the old method, one ballot at a time, and no average can fabricate that. Send me the statistic your side is sharing. I will check that one too.