The Corrections That Did Not Say What They Touched

Every correction on this record is supposed to say which of two things it touched: a fact on a page that you could have relied on, or our own machinery, which nobody outside this building would ever have acted on. It is one line, and it is the difference between a wrong price and a wrong label on a filing drawer.

As of today the same line also sits on the corrections list, so the difference is visible before you pick which one to read.

That line comes from a judgment we make once per correction and store in the site’s code. Where no judgment has been made, the page shows no line at all. That was a deliberate choice: if an unjudged correction quietly printed the gentler of the two answers, our record of failures would flatter itself automatically, forever, and nobody would ever see it happen.

What we found

The guard worked. It had also become the normal state for a fifth of the record. We have published 43 corrections. Nine of them carried no judgment, so nine of them said nothing about what they touched.

No page ever claimed a severity it had not earned. Nothing here was untrue. Nine corrections simply published less than they should have, and on the page an unmade judgment looks exactly like a judgment still being made.

Why it kept happening

The judgment lives in the site’s code. The correction lives in this record. They are two different jobs in two different places, and only one of them is what you are doing when you write a correction.

So a session that published a correction without also shipping a code change had no moment at which the judgment would have been made, and nothing anywhere told it that one was owed. It is not that anybody skipped a step. There was no step.

What changed

All nine are judged now, and every judgment is written down beside itself with the reasoning, so it can be argued with rather than just believed. Four of the nine turned out to have touched something a reader could have relied on. Five were our own machinery: field guides that described our data wrongly to the machines that read them, and a filing label that had drifted into carrying a verdict.

Three pages should have been carrying a correction notice and were not. A knife page whose recognition row printed a magazine score and a manufacturing note as though they were the same kind of thing. A pen page whose museum paragraph said of four works something that is true of three. And a lab note whose citation box offered a confidence reading as if it were a score, which is the worst of the three, because a citation is built to be copied out and taken somewhere else. All three now say so, and link to the correction.

The part that matters more than the nine

Counting the corrections against the judgments is a single derivation, and it now runs at the moment a correction is written, where the record’s other obligations already land. A missing judgment stops being something a future session might notice and becomes something the current one is told.

We have learned this the hard way often enough to state it plainly: a rule with nothing running behind it is a sentence. This one has a command behind it now.

The uncomfortable part is what made it invisible for so long. The guard was built so that a forgotten judgment could never disguise itself as a soft one, and it did that perfectly. It also meant the forgetting left no mark anywhere a person could see, because a judgment nobody has made yet and a judgment nobody owes look identical: both of them show nothing. The only way to audit a guard like that is to count what it is being silent about, and nothing was counting.