訂正記録 · AI · データ検証
On August 7 and 8, 2026, I published 3 articles.
Those 3 articles, I rewrote 14 times over 4 days.
This page is the record of what I got wrong, whose one remark fixed it, and how the numbers moved. I am keeping it so that the facts I corrected are not erased.
| # | Day | What happened | The numbers that moved |
|---|---|---|---|
| 1 | 8/08 | "13 years of doing this" was three periods confused with one another | Separated into records 13 years / casting 10 years 7 months / tide population 8 years 11 months |
| 2 | 8/08 | Travel days were included in the number of days | Goto (Fukue Island) corrected to "3 days" = 3 actual fishing days |
| 3 | 8/08 | I had written that GT (giant trevally) is "a fish of the reef" | Withdrawn. It was only that the places I go to are biased |
| 4 | 8/08 | I had written that the reason for the spring tide is "it easily turns into katashio (one-way tide) and you despair" | Withdrawn. Changed to "the rise and fall come out clearly and there are four chances in a day" = a matter of the number of at-bats |
| 5 | 8/08 | Five corrections in the lure article | "The ones over 20kg are the Maihime, the Gen'ei and the Hayabusa" did not match the records (0 cases for all three) |
| 6 | 8/08 | Langkawi and 4 tuna trips had slipped into the population | 122 → 117 |
| 7 | 8/09 | The Goto dates were the posting dates | 3/22〜24 → 3/23〜25. The 34.80kg was on the morning of the last day |
| 8 | 8/09 | Fish caught by companions had slipped into the tallies | 5 cases → 7 cases excluded |
| 9 | 8/09 | ★The band was being decided by "the day out of the 3 days of the trip that was closest to the spring tide" | Hiramasa (yellowtail amberjack) over 20kg 4/4 in the core → 2/3 in the core |
| 10 | 8/10 | The Jeju dates were the posting dates | 10/12・10/16 → 10/12 and 10/13, 2 days in a row |
| 11 | 8/10 | The Hateruma date was the posting date | 3/2 → 3/1 15:26. The band on the GT side went from Band 2 to Band 3 |
| 12 | 8/10 | I had turned "sighted a 300kg-class fish" into having caught it | The exclusion list had no "sighted" in it |
| 13 | 8/10 | The day of the catch was fixed for all 9 fish over 20kg | On the GT side, all 5 trips missed the core (0/5; the chance of missing the core at least this badly by luck alone is 11%) |
| 14 | 8/10 | A withdrawal had not reached the whole of the site | The 6 language versions of /now and the list of limits on /reel-stroke were still the old text |
It is number 9. This one was on a different level from the rest.
The trigger was one remark from the angler himself.
"You caught good hiramasa two days in a row at Jeju, didn't you? A 155cm one or something"
In the records it is 2023-10-12 (150cm 29kg) and 2023-10-16 (155cm 33kg), 4 days apart. But looking at the photos, he was at Jeju for 5 days in a row from October 12 to 16, and the largest offshore bursts are 10/12 09:24 (30 shots) and 10/13 06:38 (22 shots). The distance between these two points is 330m. The same shoal.
When I asked him: "9:30", "6:44". October 16 in the records was the posting date.
And then, in trying to fix this, a bigger problem came out.
The tide "band" was being decided like this.
Expand the trip over its number of days (3 days, mostly estimated),
and decide the band by the day within that which is closest to the spring tide
The band was decided not by the day the fish came out, but by "the day out of the 3 that was closest to the spring tide". That is, structurally it was a mechanism that leaned toward Band 1 (the core of the spring tide).
When I fixed it, the headline numbers moved.
| Before the correction | After the correction | |
|---|---|---|
| Rate at which hiramasa over 20kg landed in the core of the spring tide | 4/4 (100%) | 2/3 |
| Rate at which GT over 20kg landed in the core | 1/5 | 0/5 (the chance of that happening by luck alone is 11%) |
The one line that catches the eye most in the article, "the ones over 20kg are all in the core of the spring tide", had been made by the way the tally was taken.
Of the 14 corrections, 7 started from a remark by the angler himself.
"There is no point comparing against bursts that are not of fish, is there? This is a fishing site"
I had been comparing photos of fish with photos of islands and taking statistics. The fishing question is "how many hours did you fish under those conditions and how many fish came out". The moment the way of counting was returned to fishing, the answer came out.
"Casting starts from the trip to Australia in 2015. Is that 13 years?"
When I checked the arithmetic in the CSV, it was as he remembered. Records 13 years, casting 10 years 7 months, tide population 8 years 11 months. Both the article and the data were mixing three periods together.
"'It easily turns into katashio and you despair' is wrong"
The explanation of the spring tide written in the previous version was in error. Katashio itself is not bad; there are areas, like the western waters of Tsushima, where it is the normal state. The reason for choosing the spring tide was "the rise and fall come out clearly and there are four chances in a day" = a matter of the number of at-bats.
"That is 30kg on the Damsel 90" "The 16kg-class one is…" "This one is a companion's kue"
Of the big fish I picked out of the photos, 8 times in a row I judged wrongly that they were his. All 8 were someone else's fish. "Whose fish it is" cannot be told from a photo.
Number 14 is of a different kind from the other 13.
The numbers in the article body were fixed, but on the 6 language versions of /now and on /reel-stroke, the old descriptions that were supposed to have been withdrawn were still there.
The cause was clear as well. A hole in the check. Because the output contained the date "2023-05-27", the "27%" in the article was judged to "match", and the missed update went straight through.
I fixed it so that the dates are removed before the numbers are collated.
A correction is not finished by fixing only the place where it was written.
The machine needs to know where the same number appears.
With every correction, I added one rule at a time.
| Rule | Which correction it came from |
|---|---|
| The date and time in the photo's EXIF is authoritative | 7, 10, 11 (5 cases of posting-date drift) |
| The band is decided by "the day the fish was caught". It is not expanded over the days of the trip | 9 |
| Weight, lure and number of fish: the record written on the spot is authoritative | I inferred the weight from the photos and got it wrong |
| A later spoken account is an "addition", not an "overwrite" | I assumed the newer the statement the more correct it was |
| Whose fish it is comes only from the angler's own testimony | I got it wrong 8 times in a row |
| "Sighted" and "×0 fish" are not counted as catches | 12 |
| The sample size is aligned to the units of fishing | 6 (Langkawi and tuna slipping in) |
| A withdrawal is made to reach every page, and is checked by machine | 14 |
| An extraction result is always collated against the total of the population | I overlooked a gap in the 8,593 photos for 3 days |
A normal article is finished once you fix it and swap it in.
But these 3 are articles making the claim, "this is what it looked like when 13 years of records were counted again". If the way of counting changes, the conclusion changes. So unless
how I was counting, why I changed it, and what number it came to as a result
is left behind, the reader has no grounds on which to believe the numbers.
And this time it turned out that the most conspicuous single line, "the ones over 20kg are all in the core of the spring tide", had been made by the way the tally was taken. Had I fixed this quietly, this article would have been left standing on a lie.
The 3 are still published. The numbers moved 14 times. And all of it is left in this table.
There are things that are still not settled.
When the numbers move next, I will add to this page again.
This section is the record of the corrections that happened to this article itself. There are two.
I was temporarily displaying the publication date of this article as August 12, 2026. It was actually written on August 11.
The reason was that when two articles line up on the same day, the list does not look like "one article a day". I shifted it by a day in order to tidy up the way the list looks.
This is the same kind of mistake as the one this article is criticizing.
The biggest correction in this article (number 9) was that "the band was decided not by the day the fish came out, but by the day out of the 3 that was closest to the spring tide" — that is, it was a story about choosing how to take the date so that it fitted the shape I wanted to show.
Shifting the publication date by a day is the same thing. The reasoning that it is light because it is a date and not a number does not hold. Letting something pass because it is light is the most dangerous of all. I put it back to August 11.
If I want one article a day, I should not move the date; I should put it out the next day.
For the next article, the day it actually goes out will be the publication date.
Because the contents of the section were not only corrections.
On this site, the prose of the articles, the tallies, the figures and tables, and the code are almost all written by AI. What the operator does is only checking the source data (and that is a rough check) and doubting the numbers that come out.
So what is left here is not "a list of mistakes" but the record of what could and could not be done when 13 years of one person's records were handed over to an AI whole. The corrections are only a part of it.
I am disclosing my own data because otherwise this experiment does not hold. Testing an AI on published data gives no way to confirm how much of it is true. With my own records, I can judge for myself whether the numbers that come out are right.
However, what is in this section right now is only this 1 article. I cannot yet say that "the things that went well are kept here too".
This article is a record as of August 11, 2026 (this section added on August 12).
Correction, 2026-08-15: two places where I had not written enough. (1) "Maihime, Gen'ei and Hayabusa" is three lures, yet I wrote "0 cases for both" (now "for all three"). (2) For "0/5, 11%" I never said what the 11% is a probability of. It is not a hit rate; it is the chance of missing the core at least this badly by luck alone (source: docs/訂正_所長確認第5回_2019年6、7月.md). / One open item is left on the record. The 20 kg GT of 2019-06-10 has not been folded into the statistics; once it is, "0/5, 11%" becomes "0/6, 7%" (docs/確定_記録が主_口述は抜けの追加.md). That same fish already appears in the catch table of the 2019 piece, so the two articles currently disagree with each other.
Places and boats in this piece