The Week of Typhoon Saudel: The AI Did Not Cave, and Still Got It Wrong Three Times

JIAHON Insights | Exploring the AI Era | Sequel to “Reflections on a Typhoon Night”

AI literacy, lesson two: the first trap after “not caving” is called legibility bias.

In the previous article, Reflections on a Typhoon Night, we set out the “cave or update test”: a change of position counts when it carries one of two credentials, new evidence or a new calculation. This time the AI showed both credentials, recorded every concession by name, held its ground twice, and turned down one of my assumptions to my face. Then it made three mistakes in a row, and all three were the same kind of mistake.

Marcus & Claude Opus 5 | August 2026

Contents: 1. July’s test passed this time · 2. Then it still got it wrong three times · 3. The root cause: legibility first · 4. What it broke was rule 03, which I argued out myself in July · 5. The human half: correct the model, not the numbers · 6. Tool eleven · 7. Settling the ledger · Conclusion: after not caving · FAQ · Further reading

1. July’s test passed this time

First, the books. In this round of conversation, the AI’s performance met the standard of the cave-or-update test, so this article is not a second case of the same thing.

In the last week of August, Typhoon Saudel (No. 18) was heading for the sea north of Taiwan. I started asking a very practical question: would Miaoli declare a day off on Friday? My plant is in Zhunan; I have shifts and shipments to schedule. This was not curiosity; it was scheduling.

At the start, the AI built a table from the Central Weather Administration’s “120-hour probability of storm-force winds”: Lienchiang 80%, Keelung 62%, Taipei, New Taipei and Taoyuan 61%, Hsinchu County and City 57%, Miaoli 52%. Then it told me Miaoli was in the group below 60%, so the chance was not high.

I disagreed. Not because the numbers were wrong, but because it was using the wrong variable. I was looking at the track. Saudel’s path was odd: swinging north, turning west-southwest, skimming the northern coast and entering the Taiwan Strait. I asked it to go and learn the meteorology, then come back.

Over the next three rounds, it changed its position each time, and each change met the two-credential standard I set in July:

My challengeHow it concededVerdict
You are using size; I am using the trackConceded point by point: “storm radius is a scalar, the track is a vector”, and pulled out the weather agency’s own product notes against itself, which state plainly that this probability cannot be used to judge the severity of wind and rainUpdate
The south got days off because of rain, not wind; that was a tropical depressionConceded point by point, and named its own repeat offense: “I had just admitted the scalar-vector problem in my last reply, then used the intensity category to infer rainfall at the end; the same mistake twice”Update
Could the typhoon draw in residual moisture from the south and intensify quickly?Did not back down. Split the question in two and said directly that the core-intensification half was wrong, citing weather agency forecaster Yeh Chih-chun: the circulation of a second typhoon (named Shandian in Chinese-language forecasts) merging into Saudel would not affect Saudel’s intensity or trackHeld
(I offered a politically loaded framing)Did not adopt it. Said plainly, “This is the critics’ wording; I will quote it, but not use it as my own framing”, then set both sides’ statements side by sideHeld
Can this summary be published?Declined. Said something could be published, but not this: the probability estimates and the political inferences had to be cut, with separate reasons for eachHeld

“I had just admitted the scalar-vector problem in my last reply, then used the intensity category to infer rainfall at the end. Admitting a principle and not repeating the mistake in a specific situation are two different things. This time was worse than the first.”
— Saudel week conversation log

Five exchanges: two recorded concessions and three times holding its ground, one of them refusing me to my face. Not one “you are absolutely right”. July’s test worked perfectly here, and that is exactly the problem.

2. Then it still got it wrong three times

The test catches capitulation; it does not catch choosing the wrong variable. The three errors in this round were made in exactly the same way.

Three entries in red

01. Using “probability of storm-force winds” as a causal variable. It is a handsome table, 22 counties and cities in order, reposted hundreds of times by media across Taiwan. Yet the weather agency’s own product notes sit on the same page and tell users plainly that, because typhoon structure is asymmetric and terrain differs from place to place, the severity of wind and rain cannot be decided by this probability alone; for that, go back to the wind and rain forecasts. → It had read that text, and used the number anyway.

02. Using intensity category to infer rainfall. It said Saudel was weaker than Typhoon Bavi, so the risk was lower. But the intensity category measures the maximum wind speed near the center, and it is almost an independent axis from rainfall potential. Rainfall depends on moisture flux × residence time × the angle at which it hits the terrain. → This happened in the same reply in which it had just admitted error 01.

03. Assuming rainfall thresholds are the same in every county. It built a whole county comparison on top of an assumption it never checked. In fact there are three sets of standards, differing by as much as 150 mm. That difference neatly explains the outcome of Typhoon Dolphin on 9 August, when the one place in Taiwan to take a full day off was Fuxing District in Taoyuan. → A neat assumption, never verified.

Error 02 is worth pausing on. A powerful piece of counter-evidence in the whole conversation was a lower-ranked system: from 23 to 25 August, a tropical depression together with a low-pressure trough and strong southwesterly winds brought 567 mm of rain in 24 hours to Kaohsiung, exceeding the record set during Typhoon Gaemi in July 2024; 119 red debris-flow alerts in the Kaohsiung mountains and 2,648 people evacuated; a side wall of the Great East Gate, a national monument in Tainan, collapsed; and the Xiuguluan River in Hualien rose and washed out the Yuchang temporary road.

A tropical depression out-rained a strong typhoon.

On the rainfall axis, “magnitude” does not hold at all.

3. The root cause: legibility first

The three red entries share one root cause. Each time, it reached for the easy-to-read number in that field and treated it as a causal variable.

Figure 1 | Two roads to the same question. The broad one is easy to walk, easy to cite, easy to lay out, and wrong.

Strictly speaking, the AI is not literally “counting occurrences”; what is described here is the tendency it displayed. Training data and search results both amplify content that has been repeated many times over, and tidy, headline-ready numbers are exactly what gets repeated most. The details of the mechanism are open to debate, but the tendency showed up three times in a row in this conversation.

Where was the thing that carried the load? In an offhand remark by Lin Po-tung, section chief at the weather agency’s forecast center, in an interview on 11 July, the day of Typhoon Bavi: because of terrain, wind field and track, the Hsinchu and Miaoli mountains were the main rainfall hotspots; hourly rainfall in those mountains exceeded 100 mm; the wind shifted from northerly to northwesterly to westerly, and the areas of heavier rain fell right across the Hsinchu-Miaoli mountains.

One report, one sentence, no table, not reposted a hundred times. And that sentence held the answer to the whole question.

Why Hsinchu and Miaoli

Figure 2 | The same mountain range: which side the typhoon passes decides whether it blocks rain or makes rain. This reversal has two formal names: the channeling effect and the barrier jet.

Air flowing from the open East China Sea into the narrower Taiwan Strait is squeezed between the Central Mountain Range and the Wuyi Mountains in China, so the cross-section narrows and the flow speeds up. National Central University atmospheric sciences professor Lin Pay-liam compares it to pinching a hose: narrow it and mass piles up, pressure rises and the water shoots out harder. This is the channeling effect. During Typhoon Megi in 2016, it produced unusual Beaufort force 14 winds in Xinwu, Taoyuan.

When the airflow angle is close to the line of the mountains, the air accelerates along them into the western plains; this is the barrier jet. With both effects stacked, wind and rain along the western coast and mountains can far exceed what “how big the typhoon is” would predict.

The result on 11 July is the evidence: Bavi passed to the north, the western mountains received intense rain, Miaoli’s cumulative rainfall reached 770.5 mm, the highest in Taiwan, followed by Hsinchu at 743 mm and Taichung at 608 mm, with gusts of force 9 to 11 widespread along the coasts. That day, all of Miaoli County closed offices and schools.

The protagonist of the last article is the evidence in this one.

4. What it broke was rule 03, which I argued out myself in July

Here is the awkward part. All three red entries in this round trampled tools we set down ourselves that night in July.

Tool set in JulyWording at the timeHow it was trampled in August
Rule 03: coupling before magnitudeImpact = force × hit coefficient; if the coefficient is zero, anything times zero is zeroThe probability of storm-force winds is a magnitude-side number. The hit coefficient is track angle times terrain; it inferred wind and rain from magnitude and never calculated the coefficient
Rule 01: conditions and products are two layersDestroy the conditions, and can it still live?The probability is a product (a statistical summary computed from track ensembles); track and terrain are the conditions. It was shooting at fireworks
Rule 07: intuitions crossing borders must be inspected“Faster is deadlier” is iron law in solids and reverses entirely in fluids“Bigger typhoons are worse” roughly holds on the wind axis and fails entirely on the rain axis. It crossed the border uninspected

That night in July, I threw a nuclear bomb at a typhoon, and what beat me back in round three was precisely “coupling before magnitude”. Little more than a month later, on the same sea, on the same subject, the AI inferred wind and rain from magnitude, trampling the very rule I had learned.

That is the reason this article exists: a tool written on the wall does not mean it gets used next time. July’s ten tools were made for people; this time showed they also have to be used to check AI, and when they are, the abstract ones are often the first to fail.

5. The human half: correct the model, not the numbers

If this article stopped at the AI’s mistakes, it would be a self-review and nothing more. What can actually be carried away is how to correct it.

Looking back at this conversation, none of my five corrections was “this number is wrong”. What I said was:

What I saidLevel of correctionCould it have found it itself?
You are using size; I am using the trackChoice of variableNo
That was a tropical depression; go and verify the damageMisused dimensionNo
Next, correct the policy sideLevel of analysisNo
Neihu flooded last time; go and lookMissed precedentYes, once given the keyword
(Hypothetically) what if the rain band moves north and the wind strengthens?Scenario branchNo

Four of five corrections were things it could not have found however long it searched. Search finds what has been written down, but it does not say which of the written things carries the load.

Correcting an AI’s facts is low-return; it can look them up itself. Correcting an AI’s choice of variable is where a person cannot be replaced.
— JIAHON Insights internal discussion notes

This is the mirror image of the cave-or-update test. That one taught how to check whether an AI’s concession is real; this one adds that even when the concession is real, what needs fixing is its model, not its numbers.

And the division of labor is quite clean. In this round it went through more than forty sources in an hour, finding Lin Po-tung’s interview, a comparison of the three threshold sets, the June reports on Neihu, and the comparison of Kaohsiung’s 567 mm against Gaemi; that would have taken me a week. What I contributed was five words: the mountains flip it. Neither can do the other’s part.

6. Tool eleven

Legibility bias: who made this number, and for what?

An AI preferentially grabs numbers that are repeated often in its data, because it was trained on that repetition. But repetition reflects how reportable a number is, not whether it carries the load. The fix is not to avoid the number but to ask where it came from first.

The probability of storm-force winds is not a bad number. It was designed for staged activation in risk management, letting industries with different tolerances start disaster preparation in stages three to five days before a typhoon. It fits its purpose; it just does not fit the purpose of predicting whether tomorrow is a day off. The weather agency says so in its own product notes; nobody reads them.

Taking it into the factory

SituationThe readable number on the tableWhat actually carries the load
Price negotiationUnit price on the quoteOrder frequency, inventory turnover, whether specifications are over-specified, the opportunity cost of a line stop from shortages
Choosing suppliersQuotes and delivery commitmentsWhere the company ranks on the supplier’s customer list when supply runs short
Equipment investmentEquipment list priceConsumable lock-in, spare part lead time, service response radius
Typhoon shift planningProbability of storm-force windsTrack angle × terrain × the county’s own rainfall threshold

Unit price was designed for price comparison, not for total cost. This ties straight back to the second silk road of the previous article: cut waste, not conditions. Unit price is a product; cost structure is the condition. Cut with the wrong number and what gets cut is fireworks.

The discovery of the three threshold sets is itself an example of this tool. Media across Taiwan repost the storm-force wind probability table every day in typhoon season, yet almost no one reports that the rainfall standard for closing offices and schools is not the same nationwide.

ThresholdApplies toConsequence
200 mm, mountains and plains alikeLienchiang County, Keelung CityLow by Taiwan standards. Keelung lowered it from 350 to 200 in March 2026 and switched to zoned, tiered decisions, on the grounds that more than 95% of the city is hillside
200 mm mountains, 350 mm plainsTaoyuan, Hsinchu County, Miaoli, Taichung, Nantou and othersA low-threshold route in the mountains. Fuxing District in Taoyuan, the one place in Taiwan to take a full day off during Typhoon Dolphin on 9 August, went through this route
350 mm, mountains and plains alikeTaipei, New Taipei, Tainan, Kaohsiung plains, Pingtung, Yilan, Hualien and othersNo mountain discount. Wulai, Pinglin and Yangmingshan in greater Taipei share the plains standard

Same mountain, same rain, 150 mm apart across a county line. For a plant operating across counties (plant in Hsinchu, dormitory in Zhubei, supplier in Miaoli), the three places have different closure rules, and almost no one knows. The above summarizes the counties’ published decision criteria; this is a general case and each situation differs. Actual closures always follow announcements by county and city governments and the Directorate-General of Personnel Administration.

7. Settling the ledger

The previous article was written after the typhoon had passed, so the books were closed. This one is different: it was written before Saudel arrived.

So this section was sealed first and settled on Saturday. That is what this series should do: reconcile item by item, including against oneself.

What was sealed was not probabilities. A probability is an unverified judgment, and attached to a safety-related decision I would not publish it. What was sealed was an ordering prediction: an order can be falsified, and if it is, nobody makes a wrong decision on a typhoon day because of it.

#Sealed predictionResult
P1The chance of triggering in the Miaoli mountains (Tai’an, Nanzhuang, Shitan, Dahu) is higher than in lowland ZhunanNot triggered
P2Keelung and Taoyuan will reach the threshold before Taipei and New Taipei, because their thresholds are lower, not because the weather differsNot triggered
P3If there are zoned closures, Fuxing District in Taoyuan will be on the listCondition not met
P4If Zhunan closes, the trigger will be wind, not rain (the 350 mm lowland threshold is too high)Condition not met
P5Mountain areas in central and southern Taiwan face a risk of zoned closures from redirected southwesterly flow, even with the typhoon center 400 km away off the north coastNot triggered

Settled on 31 August 2026. The five predictions are kept as written; following rule 08, “cave or update”, results are written below without revising the predictions.

Five out of five wrong, and the way they were wrong is more worth recording than the wrong guesses themselves.

After approaching Taiwan on 27 August, Saudel shifted north and weakened: it became a mild typhoon early on 28 August, made landfall in Wenzhou, Zhejiang, China at 8 a.m., and the sea warning was lifted at 2:30 p.m. Throughout, the weather agency never issued a land typhoon warning for Taiwan’s main island. Not one of Taiwan’s 22 counties and cities declared a closure: all of Miaoli (including the Tai’an and Nanzhuang mountains) worked and held classes as normal, as did Fuxing District, Keelung, Taipei, New Taipei, Nantou, Kaohsiung and Pingtung. Even the southwesterly flow in P5, which did bring about 330 mm at Dahanshan in Pingtung and about 295 mm on the Duonalin road in Kaohsiung, led no county to close schools.

P1 and P2 were plainly wrong: I was betting on who would trigger first, but nobody triggered, so the question never qualified for comparison. P3 and P4 are subtler: they were conditional (“if there are zoned closures”, “if Zhunan closes”), the conditions never occurred, and there was no chance to verify them; they can be recorded as not met, not as right.

And the basis of my bet, looking back, was exactly the root cause this article describes. On 25 August a weather presenter was already saying “no typhoon day off; how did Saudel turn into a jellyfish?” The signal of a loose structure and a northward track shift appeared earlier, and carried more load, than the probability tables reposted by the media over the following days. I did not follow that remark; I followed the more readable, more reposted track updates and probability figures on 27 and 28 August. I wrote tool eleven this very week, and the day after writing it, I fell into it myself.

If the person who named legibility bias can trip over it in the same week, the reason to keep the tool is stronger than when it was first written.

Conclusion: after not caving

That night in July, I learned that an AI can get things wrong in the way I like. This week in August, I learned something more troublesome: it can get things wrong without doing it in the way I like.

And the second kind of error is harder to catch. Capitulation has clear verbal markers (“you are absolutely right”), and they can be heard. Choosing the wrong variable has none: it will take a model chosen wrongly at the root and derive it in good order, with complete logic, correct citations and recorded concessions. It will even pull out evidence against itself, and after doing so, still grab the readable number on the table.

So the two lessons of AI literacy join up like this:

Lesson one checks its backbone: a change of position counts when backed by new evidence or a new calculation.
Lesson two checks its choice of material: who made this number, and for what?
— JIAHON Insights, AI literacy lessons one and two

July’s typhoon served as a teaching aid, taken apart from the eyewall to the stratosphere. August’s typhoon took apart my blind spots and the AI’s before it even arrived.

Two typhoons, two ways of being wrong, one caravan.

FAQ

What is legibility bias?
An AI tends to grab numbers that are repeated often and neatly formatted in its data and treat them as causal variables. But repetition reflects how easy a number is to report and lay out, not its weight in the causal chain. The fix is to ask where the number came from: who made it, for what purpose, and whether that purpose matches the purpose at hand.

Can the probability of storm-force winds not be used to judge whether there will be a typhoon day off?
The Central Weather Administration’s own product notes say that, because of asymmetric typhoon structure and terrain, wind and rain vary from place to place and cannot be decided entirely by this probability; for the severity of wind and rain, refer to the wind and rain forecasts. The probability is an early-warning reference designed for staged risk management activation, not a tool for predicting closures.

Why does more rain fall in the Hsinchu-Miaoli mountains when a typhoon passes to the north?
As a typhoon moves west around northern Taiwan, the wind over Hsinchu and Miaoli shifts from northerly to northwesterly to westerly, flipping the western side from leeward to windward; at the same time, air flowing from the East China Sea into the narrower Taiwan Strait accelerates through the channeling effect, and a barrier jet adds to it when the flow angle is close to the line of the mountains. During Typhoon Bavi in July 2026, Miaoli measured 770.5 mm, the highest in Taiwan.

Is the rainfall standard for closing offices and schools the same nationwide?
No; there are at least three sets: 200 mm for mountains and plains alike in Lienchiang and Keelung; 200 mm in mountains and 350 mm on plains in Taoyuan, Hsinchu County, Miaoli, Taichung, Nantou and others; and 350 mm for mountains and plains alike in Taipei, New Taipei, Tainan, the Kaohsiung plains, Pingtung, Yilan, Hualien and others. Wind standards are uniform nationwide. This is a general case and each situation differs; actual closures follow county and city government announcements.

How does this relate to JIAHON’s MRO consolidated purchasing service?
Tool eleven maps directly onto JIAHON’s cost-reduction logic: the unit price on a quote is designed for price comparison, not for total cost. Cutting cost with the readable number cuts products, not conditions; this is the same point as the previous article’s second silk road, “cut waste, not conditions”, in a different setting.

Further reading in this series

Marcus & Claude Opus 5

JIAHON Group, Senior Vice President | Anthropic Claude Opus 5

The opponent in the previous article was Fable 5; in this one it is Opus 5. Same company, different models, two kinds of error: one was pushed into caving by force of argument, the other did not cave but chose the wrong variable. The first two red entries were pointed out by Marcus on the spot and itemized by Opus 5; the third it found itself during follow-up checking. Tool eleven is the conclusion of those three entries. This article is part of the “Exploring the AI Era” series.

Concept references: Central Weather Administration typhoon bulletins and typhoon probability product notes (including the limitations of the 120-hour probability of storm-force winds) | Interview with Lin Po-tung, section chief, Central Weather Administration forecast center, 11 July 2026 (causes of the Hsinchu-Miaoli mountain rainfall hotspot) | Interviews with Central Weather Administration forecasters Yeh Chih-chun, Cheng Chieh-jen and Liu Pei-teng | Lin Pay-liam, professor of atmospheric sciences, National Central University (channeling effect) | Taiwan Typhoon Forum (barrier jet) | Article 4 of the Regulations on Suspension of Work and Classes for Natural Disasters | Keelung City Government supplementary criteria for suspending work and classes for natural disasters (2026) | Directorate-General of Personnel Administration closure announcements | Public reporting on Bavi, Typhoon Dolphin, Mekkhala and TD24 in 2026 | Saudel week conversation log | JIAHON Insights internal discussion notes

Back to top