The draft was fixed. The preview still said something else.

On 11 September 2026, one of my published notes contained two different descriptions of the person it was written for.

The summary began:

For a nontechnical expert, the practical leverage of AI is a recurring job wrapped around a model — context, instructions, tools, outputs, state, logs and review — so repeated handling can move to AI while human judgment shapes later runs.

The body made a more precise distinction:

Nontechnical describes the person’s ability to build the machinery, not their ability to judge the domain.

Those sentences can both sound plausible. They do not mean the same thing.

“Nontechnical expert” makes expertise part of the audience label. The body uses “nontechnical” only to describe whether someone can build the machinery. The person might have deep domain judgment, or they might not. If they do not, the next sentence says the system needs reliable references, a capable reviewer or narrower authority.

The body therefore contained the distinction I meant. The summary still compressed it into the older phrase.

This was visible on more than one surface. In the pre-repair snapshot, the summary appeared with the article and in the site’s note listings. The same field supplied the page description and its social-sharing metadata. I did not test how a search engine or social platform would render that metadata; those services can choose different text.

The observable failure was simple: the correction reached the body and missed the summary. Why the execution stopped there is unknown. What I meant by “nontechnical” was already settled.

The published package did not show the correction history. It simply presented two versions of the argument. A reader who met the summary first had to resolve a disagreement I thought had already been settled.

On 11 September, the proposed remedy was prepared but had not yet been verified on the site: replace the old summary with one that states the article’s actual job, then inspect every rendered surface that uses it. The replacement summary was:

A useful AI conversation can still leave the next session starting over. This article follows an attempt to make a writing job resumable through files, explicit decisions and checks—and what failed when those checks existed only on paper.

The practical check is small. After correcting a meaning, do not stop at the paragraph you edited. Follow that meaning through the complete package a reader can encounter. Compare the wording and the underlying distinction, then inspect the rendered pages and emitted metadata. Historical quotations can preserve the old language when they clearly identify it as the old version.

Passing this check shows that the correction reached the relevant surfaces. It does not show that the argument is good or the claim is true. Those still require their own judgment and evidence.

Check a correction across the whole piece

Use this after changing a claim, distinction, audience description or other meaning that appears in more than one place.

1. State the correction in meaning. Keep the old wording, the intended meaning and any distinction that must survive. A writer’s reaction can be enough to start the investigation; the assistant can help locate what changed.
2. Find every affected surface. Check the title, summary or standfirst, body, listing excerpt, page description and sharing metadata. Include any generated card or download that repeats the idea.
3. Compare meaning, not only strings. A paraphrase can preserve the same mistake after the rejected phrase disappears. An old phrase may also remain legitimately inside a clearly marked historical quotation.
4. Inspect the rendered result. Confirm what the reader can actually see and what metadata the site emits. Record when and where the check ran.

Example

An article summary described the reader as a “nontechnical expert.” The corrected body said that “nontechnical” referred to the ability to build the machinery, not the person’s domain judgment. Searching for the old phrase exposed the missed summary, but the real check was whether every surface preserved the corrected distinction.

Limit

This check tests whether an intended correction reached the package. It does not establish that the correction improved the argument, that a factual claim is true, or that an external service will display the supplied metadata. Review those questions separately.