Everything here will eventually be wrong about something. This is where that goes.
What counts
A correction is anything that would change what you concluded. A wrong number, a wrong mechanism, a wrong attribution, a claim that doesn’t survive contact with better information. Typos and tightened sentences don’t make this page — they’d bury the things that matter.
Three kinds, kept separate on purpose:
Correction — a fact was wrong.
Retraction — the claim doesn’t survive and I’m withdrawing it.
Update — something changed after publication. The original wasn’t wrong; the world moved.
Conflating “I was wrong” with “things changed” is how being wrong gets laundered. I’d rather not have the option.
How it works
When I find one — or you do — two things happen. It goes here, and a banner goes on the piece itself, naming what was wrong and quoting the original line.
Both, not either. A corrected version that only announces itself inside the corrected version is no help to whoever opens the old one under pressure, and it is always under pressure.
The wrong version stays legible. I don’t quietly edit and move on.
Found one?
patrick@pi0.co
I would rather be corrected than right, and it isn’t close. If you’re the reason something on this page exists, I’ll name you unless you’d rather I didn’t.
The log
16 September 2026 — What’s in a Name? Everything — to one model, and nothing to the other · Update
What it said: “Strip the vocabulary noise, the misclassifications, and the fragments of split entity names, and the honest number is 5.9%.” Published 19 August 2026, with a table giving cross-matter linkage on genuine identities only as 5.9%, across 300 document pairs.
What changed: The engine did. Fixes that landed within about a day of publication changed how entities are classified, and genuine-identity cross-matter linkage is now 4 of 254 pairs — 1.6%.
Why this is an Update and not a Correction: nothing published on 19 August was wrong. The measurement was honest and the method held. The thing being measured moved. Quoting 5.9% today would be wrong, which is why this entry exists.
How it surfaced: Re-running the measurement while drafting a later piece, instead of quoting the published number.
16 September 2026 — I Owned My First Computer Outright in 1986. I’ve Been Renting Ever Since. · Correction
What it said: “The write-up, including the correction after a second provider returned zero refusals, is here.”
What’s true: That write-up carries no correction. The second-provider result — 0 refusals of 40 keyed documents on GPT-4o — was part of the original reporting on 19 August 2026, not a later correction to it. The piece does carry a “Limits” section and an “artifact we nearly published” note, both candid, but a reader following that link expecting a correction will not find one.
How it surfaced: Opening the linked piece and checking it directly, while building this page.