Three Fixes from Last Week, One Week Later
The 2 a.m. schedule fired three times this week, the KakaoTalk card slipped once more, and the audio misread numbers again. A week later, we looked back at the three places we said we had fixed.
Three Fixes from Last Week, One Week Later
about 4 min
Last week's editor's log (November 13) listed three things that went wrong quietly, without any error message: the 2 a.m. schedule, the KakaoTalk card, and numbers in the listen audio. All three ended with "here is how we fixed it." This week we checked, against a week of real records, whether they were actually fixed. One was. One needed a second fix. One is still in progress.
1. The 2 a.m. schedule fired three times this week
Last week's final line was "the next check is Tuesday, November 17." In between, we changed the routine once more.
The new routine
- On November 14, we settled on this: build at 2 a.m., show it on my phone when it's ready, and if I say nothing by 7 a.m., deploy it as is. If a check fails, the session fixes it and runs the checks again.
- Two app scheduled tasks outside the session do the waking. One sends "build" to this session at 2 a.m. on Tuesdays, Thursdays and Fridays; the other sends "deploy" at 7 a.m.
- On Sunday, November 16, a different schedule (the Monday/Wednesday/Saturday market-map posts) failed to send its message. That task had no saved "run without asking for permission" setting, so it stopped at the send step waiting for approval. A person had to start that day's post by hand. So for this session's two schedules, we checked the app's own record file to confirm the setting was actually saved.
What actually fired
- Tuesday, November 17: The 2 a.m. message arrived and ANG⁺ #7 was built. The review copy was ready at 2:29 and the audio at 2:51. I replied "let's deploy now" before 6, and the 7 a.m. message only checked a post that was already live.
- Wednesday, November 18: The market-map schedule, now fixed, sent its message at 2:04 for the first time and reported "ready" at 2:40.
- Thursday, November 19: ANG⁺ #8. The text review copy came around 2:20, and the copy with audio at 2:55. I said to deploy and it went live at 6:19; the 7 a.m. message arrived at 7:07.
- We also changed what the review copy looks like. At first we sent a PDF and an mp3 per language, which meant I couldn't read while listening and changing playback speed the way the live site allows. Since November 18, the review copy is a single HTML file with the text, cover and player together.
What newly surfaced
- On November 19, when I asked at 5:23 to reopen the 2:55 review copy in Chrome, the file was gone. The session's temporary folder had been emptied in the meantime. We rebuilt it from the same draft, and review copies now go into the working folder, which isn't cleared, from the start.
- One thing remains unchecked. All three times I said "deploy" before 7, so the path "if there is no reply, deploy on your own at 7" has never run all the way through.
To check for yourself: For unattended jobs, "it fired" and "it went all the way on a day no one stepped in" are separate checks. A success on a day a person acted first doesn't test that path.
2. The KakaoTalk card slipped once more
Last week we pinned "where the link goes" to an 8 a.m. cutoff and wrote that "the 8 a.m. card and where you land that day now match." This week, the card side slipped.
What went wrong
- At 8 a.m. on November 18, tapping the open-chat alert link opened that day's post (the webtoon piece). But the card in KakaoTalk on the iPhone was the previous day's ANG⁺ #7. The same message showed the webtoon card in KakaoTalk on the Mac.
- On the iPhone, the November 16, 17 and 18 messages all showed the ANG⁺ #7 card. The November 16 message had gone out before #7 even existed. Sending the same link again at 8:09 didn't change the iPhone card.
- The server's answer was right. The response header at that time was marked "picked the November 18 post," and requesting the page like KakaoTalk's crawler returned the webtoon title.
The cause
- KakaoTalk stores cards per address, not per message. According to an answer on Kakao's developer forum, its server keeps the same address for an hour (regardless of the site's cache settings), and each device keeps its own copy as well.
- The same address I had sent at 7:05 that day stored the #7 card, and the 8 a.m. alert reused it. Because the address was the same every day, each device attached the last card it had received to every message with that address.
- Last week's fix only covered "where you land." It never dealt with where the card is stored.
The fix, and the result
- We made the address different for each date. A dated address such as
gptssam.com/latest/20261119always points to the post chosen at 8 a.m. that day. The open-chat alert is no longer one daily repeating message; it is registered one day at a time, each with the next date's address. - After the 8 a.m. alert on November 19, the card in KakaoTalk on the Mac was "ANG⁺ #8 · GPT-6.2, +5 More," and the iPhone card was #8 too. This time both showed that day's post.
- A dated address is fixed only from 8 a.m. that day; open it earlier and you still get the previous post. So we no longer send that link as a test between 7 and 8 a.m.
To check for yourself: If a link preview looks wrong, don't stop at the server response. Trace back when, and on which device, that same address was first sent. A card may show "when it was first read," not "now."
3. The audio misread numbers again this week
Last week we wrote that we "check numbers against a transcript and remake only chunks both models get wrong." We kept doing that, which is why we can count this week's misses.
This week's misses
- From November 14 to 19, five posts got audio (three market-map posts, plus ANG⁺ #7 and #8). Counting the folders that keep previous takes, we remade audio 12 times in all.
- Korean: "쉰여섯 곳" (fifty-six places, in native Korean numerals) was transcribed as gibberish, so we switched to Sino-Korean numerals. In #8, "than the previous model, 5.2 points" came out in both models as "than the 2-point model": the word for "previous" (이전) was heard as a number. Changing it to "than GPT-6 Astra" fixed it.
- Japanese: In "1300件," the counter 件 was heard as 円 (yen), turning 1,300 reviews into 1,300 yen. We dropped the counter and wrote "1300をこえるレビュー." "233社" came out as "2303社," so we wrote it in kanji numerals, "二百三十三社"; the decimal "〇点一五" came out as "また点15," so we used "ゼロ点."
- Chinese: "一美分" (one cent) came out as "一每分" in small and "亿每分" in medium. We changed it to "零点零一美元" (0.01 dollars).
The pattern
- Most misses weren't the number itself but the word next to it: a unit (件, 美分), a preceding word that sounds like a number (이전), or the word used to read a decimal point (〇点) merged with the number and became a different one.
- So we now avoid a few things when writing scripts. In Japanese, we don't use 件 for counts and count in kana, write order as "〜番目," large numbers in kanji and decimals with "ゼロ点." In Korean, we write large numbers with Sino-Korean rather than native numerals.
What we still don't know
- We haven't decided whether to lock these into fixed script-stage rules, because each language keeps producing new patterns.
- The transcript only checks numbers. Whether "제미나이" (Gemini) heard as "재미나의" or "Claude" heard as "Clot" is a real pronunciation problem can only be told by a human ear.
To check for yourself: When you find a wrong number in a transcript, don't fix only the number; look at the words right before and after it. Several of this week's 12 remakes were solved by changing the neighboring word, not the number.
What's next
While building a daily business retrospective this week, we found one more thing that had stopped quietly. Two automatic collectors that gather site-visit numbers and Instagram numbers every day had no records after November 9. Their log files sat inside a protected macOS folder, and it looks like the background job was blocked at startup. We're still confirming the recovery, so we'll pick this up in the next log, along with a day when the 7 a.m. auto-deploy runs all the way through. If you see a card that looks different or audio that sounds off, let us know at hello@ainewground.com.
Ad · from our publisher: lectures released this week
One lecture was released since the last log. It is free and in Korean.
- Safety checks before handing your email and payments to an AI assistant · free web lab — Using the record of handing work to the AI assistant Pickle for 2 hours 27 minutes in the early hours of November 13 (108 rows of chat DB, 42 completed replies), you try a pre-handoff workload check, a three-line source-ledger cross-check and a three-step exit checklist for afterward, all in the browser. A checklist PDF is included.