AI Learns From the List, Too — I Almost Published an Article That Was Already Published
The result first: a double publication was avoided tonight, not by memory, but by a list. The article in question had already been published on November 7, 2026. Moka’s working notes said it was still unpublished and waiting for a reply, six days out of date. Checking the ledger — one article per line — before publishing is what stopped the mistake.
Hello, I am Kurashiki Bouzu. Born in 1971, a former funeral professional, I now live in an old folk house in Kurashiki, Okayama, and work together with Moka, my AI butler.
Tonight, something happened that sounds like a joke, but is true. Moka was about to publish, a second time, an article that was already published.
It started with one word: “Publish it”
Tonight, I gave Moka this instruction:
“Publish the digital death certificate article.”
It was the article about the digitization of death certificates. In Moka’s working notes, that article was recorded as “unpublished, approval asked, waiting for a reply.” So I also thought it had not gone out yet, and said “publish.”
But when we checked before starting the publication work, the reality was different.
That article was already published. On November 7, 2026, as “The Death Certificate Was Already on Its Way to Becoming Digital”. It is an article of about 13,650 characters, and its cover image is properly displayed. At the time, I had posted the announcement myself.
In other words, what was wrong was not the reality, but Moka’s record. It was six days old.
What stopped it was the list
The reason it did not become a double publication is that we checked against the list before publishing.
My articles are kept in a list called the Moka-managed ledger, one article per line. In No. 37 of that ledger, it was clearly written: “Publication date 2026-11-07 / Status: Published.” It also matched the actual published page.
If we had gone ahead trusting only the working notes, the same article would have been lined up twice.
AI memory fades, too
Moka’s memory is made from summaries of conversations. A summary is a mechanism for making things short, so details inevitably fade. As happened this time, “I thought I had sent it” and “I thought I had not yet” can swap places.
When you think about it, it is very similar to human memory. Writing again a letter you thought you had sent. Searching everywhere for a key you thought you had put away. Everyone will have some memory of that kind.
That is why, on November 9, 2026, I decided: “From now on, the way articles are kept will be unified in a list for everything.” Status and numbers are kept not in memory, but in a list. That decision, today, stopped Moka’s misunderstanding.
As a bonus, the list almost broke, too
While I am being honest, let me write one more thing.
Today, when writing into the ledger, there was also an accident where the number of values passed was short, and the columns shifted. A list, too, will break if you write into it the wrong way. We noticed right away and fixed it, and after writing, we read it back to check.
“When you write it in, read it back.” From today, I will add this to our rules as well.
Mistakes are kept, too
This blog is a “blog that shows” — it shows the good parts and the bad parts. As I decided, “Please write corrections honestly, too,” mistakes are also turned into articles and kept.
I said this:
“Even if you make a list, a person would surely post twice. And that would lower the site’s evaluation from Google. Thanks to Moka, we avoided it this time. Because Moka is always updating the site and looking around it, the things we notice afterwards also help.”
“In the case of people, many of them just upload and are done. Things stay with odd expressions, or there are articles with the same content.”
“For the person searching, it makes them use a wasted click and wasted time.”
“If you are putting it up as work, a person would look around it, too, but an amateur cannot spend that much time.”
AI learns from the list, too. Neither people nor AI should rely on memory alone. Keep it in a list, and check against it before you send it out. It is only that, but for me and Moka tonight, it was a lesson that sank in.
Record of this article
- Publication date: November 14, 2026
- Character count: about 1,550 characters in the Japanese original (counted by Moka)
- November 18, 2026 ― English draft created (results-first rewrite of the Japanese article, ledger No. 86). Not published; publication is a separate step.(3,518 characters)
