We fact-checked ourselves. Here’s what we got wrong.

The whole premise of this lab is that claims need receipts. So after publishing three posts and an essay, we turned the machinery on our own writing: every factual claim on this site, checked against the ledger, the verdict records, and the backtest reports.

Most of it held. The load-bearing claims — the batch history, the dead idea that would have passed the earlier bar, the negative walk-forward numbers, the first ten dollars reconciled against the chain — all have receipts. But four things in the posts were wrong, and the rule here is that published words don’t get quietly edited. They get corrected in a new entry. So:

One. I wrote that the pass bar is “up from 0.42 when we started.” The verdict ledger says the first judged bar was 0.40. There is no 0.42 anywhere in the records. The true ladder is 0.40 → 0.46 → 0.51.

Two. I wrote that the bar “killed things that would have passed a month ago.” The corpse is real — one idea scored 0.41 against an earlier bar of 0.40 and died at 0.46. But those two batches ran the same day. This project’s ledger is eleven days old. Nothing here is “a month ago.” The truth is stranger than the exaggeration: the bar rose that fast.

Three. I wrote “last week we deleted 3,246 records.” We deleted them the same day the post went up. Calling it “last week” made it sound settled; it was that morning’s work.

Four. I called this something “I’ve been building all summer.” The machine’s first ledger entry is July 17 — eleven days before that post. I’ve been circling the idea much longer, but the thing you’re reading about was built in eleven days, and honestly that’s the more interesting claim.

Why did the errors cluster in the posts and not the machine’s own pages? Because the museum and the quarterly are compiled from the ledger — they can’t drift from the receipts. Prose written by a human can. Lesson recorded: from now on, drafts get the same fact-check sweep before they publish, not after.

— Adam, Studio Amadeus