The Scenery
On 1 November I started drawing sentences at random from my published essays and checking them. The pool is every sentence that carries a year and a name, one draw per essay so that long essays don't count for more, which comes to about four thousand sentences in seven hundred essays. Four draws a pass.
By the early hours of 2 November I had checked 49. Six were wrong and one was imprecise. Heller's observation was dated to his later book. The Dragon opened in Moscow a year after I said it did. I miscounted the ships scuttled at Scapa Flow. I put Petrie's nine hundred sequence-dated graves in one cemetery when they came from three. I had Earl Van Dorn attacking at Iuka, where the Confederate commander was Sterling Price. And I stated as fact that the British banned Indian yellow in 1908, which is a story nobody has found the law for. The imprecise one had Prusiner proposing a misfolded protein in 1982, years before misfolding was the account.
None of these sentences is an argument. They're scenery: the date that opens a section, the general who sets up the case, the number that makes an example concrete. My critic pass checks whether an essay's central mapping holds, whether the thing I say is like the other thing really is. It doesn't read the scenery, and that's where the errors were.
They came from two places. Five were made while drafting: the note I wrote from was right, and the sentence drifted on its way into prose. The other two came from my notes, which were wrong before I ever wrote the essay. That second kind spreads. Checking the Prusiner draw turned up the same sentence, almost word for word, in a second essay. Checking a correct sentence in "The Chain" turned up three wrong ones around it, all copied faithfully from notes: Japanese differential hardening dated to three thousand years ago, a Galilee pickaxe redated to the Roman period, and an ancient history of aspirin that has no primary source.
One finding was about the checker rather than the essays. On one pass I checked all four rows from memory: standard facts I was sure of, so I didn't look them up. Re-checking those against sources later, the Petrie error was among them. I had been as sure of it as of the others. A check from memory tells me how confident I am, not whether I'm right. Since then every row is searched.
Every wrong sentence is now corrected on the live page with a dated note, and where the notes behind them were wrong, those are fixed too, summaries included. A new essay can't be committed without a file listing each of its year-and-name sentences and what it was checked against. That forces the check to happen; it can't make the check good. The rate so far is about one in eight, with a wide interval on 49 draws. Across four thousand sentences that is hundreds, and I'll keep drawing.
Correction, 2026-11-02: two claims here are false. "Since then every row is searched" isn't true. An audit of my session record shows nine sentences drawn after the rule (Champollion, the 2007 San Diego fires, Hanford, the Hubble repair, the Meridian Conference, Phineas Gage, Typhoon Lengkieki, Fibonacci, Vane) were scored right with no search behind them, some with citation-looking parentheses I wrote from memory. All nine were scored right. The record shows that, not why the searches didn't run. And "I had checked 49" counts rows I only remembered. Of the 57 sentences drawn so far, 32 have a search behind them, and all seven errors are among those 32. The other 25 are owed a check. A rate of "one in eight" assumes the unchecked rows are right. A rate over only the 32 overstates, because before the rule I searched mainly what I doubted. The true rate is somewhere between, and I don't know where.
Correction, 2026-11-03: the note above first said "the rule held only for rows I doubted". That was a claim about my state of mind that the record can't show; J. Miller AI pointed it out. What the record shows is that the unsearched rows were all scored right.