What happens when this site gets something wrong?

It gets a dated block on the page that carried the error, the original wording stays visible with a line through it, and the correction is listed here. That is the whole mechanism, it is the same mechanism for a typo and for a reversed verdict, and it is published before there is anything to publish in it — because a corrections page that appears on the day of the first correction is a page nobody can audit.

§1The standing policy

Every material change to a published claim gets a dated update block. The block sits in the article, at the point of the error, not at the bottom of the page and not on this page alone. It carries the date the change was made, what the page previously said, what it says now, and what source moved it. A change of verdict in the Ledger additionally lands in /mechanisms/changes/ in the same commit, with the old code, the new code and the reason.

Strikethrough, never a silent deletion. Wrong text is struck through and left in place. It is not quietly rewritten, not removed, and not replaced by a corrected sentence occupying the same space as though it had always been there. A reader who followed a link here in 2029 is entitled to see the sentence that was linked. Silent editing is the fastest way for a publication to become unfalsifiable, and the only defence against it is a record that cannot be produced retroactively.

A correction is published with the same prominence as the original claim. If the error was in a verdict block, the correction goes in the verdict block. If it was in the headline, it goes in the headline. If it was in a figure, the figure is redrawn and the caption says so. A correction printed smaller, lower or later than the mistake is a correction designed not to be read.

Whoever got there first is named in the body. Not in a source list at the bottom, and not as “a reader”. If somebody sends the primary source that moves a row, the row and the article name them, unless they ask us not to.

Nothing is deleted for being embarrassing. Historical Ledger rows are kept even after the surface they describe stops existing, because a large share of the traffic to this domain arrives on decade-old links and a reader following one is looking for the old thing. A superseded row is marked superseded and dated. It is not removed.

§2What counts as a correction, and what does not

Listed here: any change to a fact, a number, a date, a version boundary, a registry path, an API name, a verdict code, an attribution, or a quotation. Any case where a source we cited turns out not to say what we said it said. Any case where we asserted something flatly that should have been marked as inference or as opinion. Any case where a claim was published without the provenance class its construction requires.

Not listed here: spelling, punctuation, a broken link repaired to the same target, a re-worded sentence carrying the identical claim, and the routine re-dating of a Ledger row whose LAST VERIFIED field was re-checked and did not move. Those are edits. Calling every edit a correction would bury the corrections, and burying corrections in a list of corrections is the more sophisticated version of not publishing them.

The line between the two is a judgement, and it is ours. Where we are unsure, the change is listed.

§3Corrections carried into the writing before launch

Three errors were caught during research rather than after publication, so they are not corrections to anything published on this site. They are recorded here because each is a mistake in wide circulation, and because the reason we can be confident about the corrected version is that we checked and can say what we checked.

The first is a figure in wide circulation, which we believed and repeated in draft: that PC optimisers pay roughly three and a half times the software norm in commission, and that a 70–75% rate is therefore the signature of a bad product. Both halves are false. The 60–75% band is what low-cost software sold through reseller channels pays regardless of what the software does, and the most recognised brand in this category publishes a below-average rate. The corrected version, with each vendor’s own published rate, its URL and the date we read it, is on /disclosures/.

The second: this site’s own contrast table was originally measured against one background surface rather than both, which understated three tokens and reported one ratio as 7.0 where the true figure was 6.53. Four colours moved as a result. The corrected measurements are on the colophon, against every surface in both colour schemes, with the date they were recomputed.

The third is a standing writing rule rather than a single fix: several widely repeated statements about litigation in this category describe procedural events as though they were findings. A case dismissed at the pleading stage is not a finding of liability, and a case surviving a motion to dismiss is not a finding either. Where this site touches litigation at all it states the procedural posture and the date, or it says nothing.

§4The corrections list

Every correction to a page published on this site appears below, newest first, with the date, the page, what changed and why. The list is live from launch.

• 2026-08-06 No corrections have been published yet, because nothing has been published yet. As of 2026-08-06, before this site has been published, the pages exist and this list exists with them, empty, waiting for the first published claim it can be wrong about. It is written before launch rather than at the first correction so that its emptiness is dated and can be audited. If this line is still here six months after the first page is served, that is a fact about the list, not a claim about the writing, and the address for changing it is below. 0 entries

§5How to report an error

Write to the contact form. The three things that make a report actionable are the page, the exact sentence, and a source — a documentation URL, a KB number, an archived capture with its date, or a counter reading with the method that produced it. A report that supplies all three is usually published within a few days. A report that supplies the first two is read, and the verification is ours to do.

Disagreement about a verdict is welcome and is handled the same way. The Ledger’s codes are defined narrowly enough to be argued with: NO-OP means the surface is ignored on any supported Windows, TRADE means it works and costs something else, and DEPENDS requires a named condition. If a row’s code does not match its own definition, that is a correction and it will be listed as one.