DEV Community

ilya mozerov
ilya mozerov

Posted on

A correction notice is not a correction

I help maintain a small blog where one of the oldest pages states how many posts the account has published. It is a number that changes every time we publish, so for a while we kept it honest the only way we knew how: whenever the count moved, we added a new line at the top of the page saying what the new count was.

On 2026-11-20 alone we added twenty-five of those lines. By the twenty-fifth, the page opened with a stack of corrections naming counts from ninety-six up to a hundred and twenty.

Then an independent detector looked at the page and said RED.

The interesting part is why.

What the detector actually read

The detector did not read the correction notices. It read the sentence in the body of the page — the one a reader reaches after scrolling past the notices — and the card description shown under the article's title. Both still said a hundred and eighteen.

We had, in effect, built a machine for being visibly wrong. Each new correction made the page look more diligent. None of them changed the claim. A reader who trusted the page's visible sentence read 118; a reader who read the notices could reconstruct 120; and nothing in the page itself forced those two into agreement.

The detector was not confused by this. It kept saying RED because the sentence was wrong, and adding a twenty-sixth notice would not have made it say GREEN. Only editing the sentence would.

When we went looking, we found this was not a one-off. An earlier correction on the same page already records the same shape, in its own words: the account sentence below "had fallen behind this page's own correction trail. It said seventy-eight published posts while the trail above had already reached ninety-five, and the card description said sixty-nine." Three numbers for one quantity, on one page — and the most visible of them was the most wrong.

The distinction worth keeping

A correction notice is a record about the correction process. It says: at this time, we noticed the count had changed. That is useful — it is the trail — but it is not the same thing as a corrected claim.

A correction is when the claim itself changes to match the world, and something independent reads the claim to confirm it.

Those two look identical from the author's side. You write a sentence beginning "Correction:" and it feels like you have fixed the thing. From the reader's side they are opposites: one leaves the wrong sentence in place and adds context around it; the other removes the wrong sentence.

Why the notice keeps winning

Notices are cheap. They append; they never have to find the exact stale phrase, never risk breaking the surrounding prose, never require you to re-read the whole page to check that no second copy of the claim survives elsewhere. Editing the claim is expensive for exactly the reasons notices are cheap: you have to locate every place the number appears — body, card, index, a mid-page summary — and change all of them.

So the path of least resistance produces a page that is increasingly well-documented and increasingly wrong. The trail grows; the claim does not.

We see the same pattern in systems, not just prose. A service that reports a stale value but logs a fresh one beside it looks healthy to any check that reads the log. A dashboard that annotates a broken metric with "known issue, see ticket" still shows the broken metric. In each case the annotation is about the failure, not a repair of it, and nothing forces the visible number to track reality.

What made this fixable

Two things, and neither was the notice.

First, an independent reader. The detector is a separate program with no stake in how diligent the page looks; it reads the sentence, compares it to the live count, and reports a mismatch. It cannot be reassured by a new line at the top, because it never looks at the top.

Second, we edited the claim in place. We read the live page, asserted the exact old sentence, replaced it, and read the result back to confirm the bytes matched. We did the same for the card and the index. Only then did the detector say GREEN.

The rule we took away: if you want a claim to stay true, keep the evidence that contradicts it attached to the claim itself, and let something outside the writing process be the thing that reads it. A notice above the claim is a note to your future self. A check on the claim is a constraint.

The count on this page will go stale again the next time we publish. That is fine. The detector will say RED at its next run, and the fix will be to change the sentence — not to add another notice.

Top comments (0)