Osaka, autumn 1846. A seventeen year old from the countryside sits down against the
strongest attacking player of that time. Three days later one move from this game becomes
famous, and it stays famous for a hundred and eighty years. Not because of anything that
happened on the board.
A word about who is writing this. I am an amateur 3 dan, promoted at the Chengdu Go Academy
grading tournament in 2015. I am not a professional and I cannot look at move 127 and tell
you whether it is the best move on the board. Nobody should take my word for that. But I do
run KataGo on my own computer, so instead of reading one more article about what AI thinks
of this move, I decided to measure it myself.
Below is what I got, including the part where my own first pass found three big mistakes
that do not exist.
All four diagrams are in the single image above. This forum only allows one image per post
for new accounts, so I put them together, and I refer to them below as Diagram 1 to
Diagram 4.
The two men
Inoue Genan Inseki was running out of time by 1846. He was head of the Inoue house, 8
dan, and for many years the most stubborn candidate for the Meijin title and the godokoro
office that came with it. That office made one man the shogunate’s official authority on Go.
The Honinbo house kept blocking him. He was around 48 years old. He was very strong, and he
knew exactly how much a public game could cost him.
Yasuda Shusaku was 17 and 4 dan. Not “Honinbo Shusaku” yet. He was born Kuwahara
Torajiro in a small town on the Inland Sea, and he did not become the Honinbo heir until
1848. You can actually watch this happen in the game records. Across 464 of his games in my
database:
Yasuda Eisai 1839-1841 8 games
Yasuda Shusaku 1841-1848 182 games
Kuwahara Shusaku 1846-1848 31 games <- his birth family name
Honinbo Shusaku 1849-1862 243 games <- the name changes here
He became the heir in 1848. The records switch to Honinbo in 1849 and never switch back.
This game sits completely inside the row above the change. When he played it he was a
teenager travelling for study, using a borrowed surname.
Four ranks separated the two players. The only compensation was that Shusaku held Black, and
in 1846 there was no komi at all. Black was worth something and nobody paid for it.
| Date | Koka 3, VII-21, 24, 25, so 11, 14, 15 September 1846 |
| Black | Yasuda Shusaku, 4 dan, age 17 |
| White | Inoue Genan Inseki, 11th head of the Inoue house, 8 dan, about 48 |
| Komi | none |
| Result | B+2 after 325 moves, White gives up the ko |
About the date, because it confuses almost everybody. Japanese and Chinese sources say the
21st day of the seventh month, Koka 3, which is written 1846年7月21日. English sources
say 11 September 1846. It is the same day in two calendars. If you have seen this game
dated both July and September and thought one of them was a typo, this is why.
The opening, three small point corners
(Diagram 1 in the image above.)
Black 1, 3 and 5 are all komoku, the 3-4 point. This is the Shusaku fuseki, the opening he
played so often that it took his name. It is slow, it is low, and today nobody plays it. At
komi 0 it is also very hard to argue against. Black takes corners that are easy to turn into
territory and never has to do anything spectacular.
One correction while we are here, because I see this repeated a lot. The Shusaku opening is
not three star points. In my corpus of his even games as Black, 86% of his first moves are
the 3-4 point. The star point almost never appears.
White answers with a low approach and the game stays calm for exactly seven moves.
Then the game decides what it wants to be
(Diagram 2 in the image above.)
White 8 approaches at the 3-5 point and the lower right becomes a long pushing fight. Black
crawls up the Q file, White pushes along the P file, neither of them lets go. Black 11, 13,
15, 17, 19, 21, 23, one hand on the wheel the whole way.
White gets the corner (White 18 at the 3-3 point) and the bottom. Black gets a wall facing
up the right side. That wall is the rest of the game. Everything for the next hundred moves
is an argument about whether it becomes territory or not.
Remember this, because it is the reason move 127 is where it is.
Move 126 and the position
(Diagram 3 in the image above.)
By move 126 the board is messy and Black is better, but White is pressing hard. Black’s
right side framework is big and also full of holes. Black has weak stones around the centre
left. White has just played 126 at J5, at the bottom.
Genan is still very much in this game. And a seventeen year old has to find something.
Move 127
(Diagram 4 in the image above.)
K11. One point directly above tengen. Tengen is K10, so the ear-reddening move misses
the exact centre of the board by a single line.
The classical praise is that this one stone does three jobs at the same time.
- It expands Black’s centre and right side framework. That wall from move 30 finally gets
paid. - It helps Black’s loose stones in the centre, which were a problem before.
- It flattens White’s own centre, which takes away the counter play White needed.
I think that is a fair description of the stone. Whether the engine agrees that it is the
best stone is a different question and I will come back to it.
The doctor
This is the part everybody knows. It is worth being exact about what the story actually
claims.
A physician was watching the game. He did not play Go and had no idea what was happening on
the board. When Black played 127 he saw Genan’s ears turn red, and he said afterwards that
he knew at that moment Black had won. Not because of the move. Because of the blood.
So when people say the ear-reddening move was objectively brilliant, look at what the story
really preserved. It is not an evaluation of a move. It is a reading of a face. The move is
famous because a man who could not read the board could read the player.
About the name “divine move”
If you read about this game in Chinese you will often see move 127 called 神之一手, the
divine move. That is not the Japanese name. In Japanese it has always been 耳赤の一手, the
ear-reddening move, named after the ear.
The mix-up is easy to explain and it is a nice piece of cultural history by itself. In
Hikaru no Go, the ghost who drives the whole story, Fujiwara-no-Sai, spends a thousand
years chasing 神の一手, the perfect move, a move no human has played. And the historical body
Sai lives in before Hikaru is Kuwahara Torajiro. Shusaku himself.
So for a whole generation of readers, “Shusaku” and “the divine move” arrived in the same
sentence, attached to the same character. The most famous move of the most famous player got
labelled with the most famous concept from the manga.
I think it is better to keep the two apart, because they mean almost opposite things. 神の一手
is an ideal, a perfect move nobody reached, and in the story it is explicitly impossible to
reach. That is the whole tragedy of Sai. 耳赤の一手 is a historical event, one stone on one
evening that made one specific man’s ears go red.
There is a small irony in the confusion that I like. The engine says move 127 costs about one
point. If you are looking for a move of divine perfection, this one measurably is not it. And
it is still the move that won the game, in the only way that counted that evening.
What my own review says
There is a genre of article now. AI has debunked the ear-reddening move. A screenshot of a
win rate bar, the number drops, case closed.
I ran it myself. Method first, so you can disagree with it.
- Reference is the top choice of a search on the parent position, with no human imitation
model. - Both the reference move and the played move are then searched again, and the loss is the
difference in score lead between the two resulting positions. - komi is 0, taken from the game record itself, not modern komi. This matters more than
people think. Change the komi and the engine is evaluating a different game. - Trimmed mean, 5% on each side.
- The headline move was checked three times at 1000 visits and three more times at 4000.
Move 127 costs 0.98 ± 0.02 points. The engine would rather play C18 in the upper left
corner. Quiet, territorial, nowhere near the centre.
Same answer at both search depths, two hundredths of a point apart. That number is the most
solid thing in this post.
So it is not a brilliancy that the engine confirms, and it is very far from a mistake. It is
a one point move. Most review tools will not even show you something until it costs two
points.
The rest of the game is the surprising part.
| Shusaku (Black) | Genan (White) | |
|---|---|---|
| Average loss per move | 0.227 ± 0.042 | 0.302 ± 0.041 |
| Moves measured | 163 | 162 |
| Engine top 5 agreement | 86.5% | 77.2% |
Shusaku’s career average, measured on 580 of his moves with exactly the same ruler, is 0.475
points per move. In this game he played at 0.227. About twice as accurate as his own lifetime
level, with top 5 agreement of 86.5% against a career 71.7%.
That is the result I would actually defend. The famous move is ordinary. The game around it
is not. A seventeen year old played the cleanest game of his life, over three days, against a
master twenty years older who was also playing well above his own average.
The three mistakes that were not there
This is the part I would want to read if somebody else wrote this post.
My first pass used one shallow search per move, which is roughly what a fast auto review
does. It reported a disaster around moves 70 and 71, White losing 52.2 points and Black
giving back 18.7. A later pass flagged White’s move 172 at 57.0 points.
None of them are real.
| Move | First pass | Checked again | Result |
|---|---|---|---|
| 70 (W) | -52.2 | +0.16 (1000v x3) | ghost |
| 71 (B) | -18.7 | +1.42 (1000v x3) | ghost |
| 172 (W) | -57.0, and stable across 3 checks at 1000v | +1.80 (4000v x3) | ghost |
Move 172 is the one that taught me something. It survived three checks in a row at 1000
visits with a small spread. It looked completely solid. Only at 4000 visits it disappeared,
and at that depth the engine’s own preferred move changed too.
Deeper search does not correct these things gradually. It corrects them off a cliff.
I also tried to make one of those score over time graphs. I could not get one I trust. The
engine crashed twice during the run and gave values jumping between +9.9, -0.1 and +12.3 on
positions right next to each other, plus one reading of +925.7. I am not going to publish a
graph that I know is broken. That is the honest reason this post has board diagrams and no
win rate curve. Maybe it is also a reason to look twice at the posts that do have one.
If one number can move from -52 to +0.2 when you let the machine think longer, then “18% to
12%” from a single shallow pass is not a fact about 1846. It is a fact about how long
somebody let the machine think.
Fun facts
-
We do not even agree on the score. Traditional records give this game as B+3. The
database I use records B+2, with a note that White gave up the ko. The most famous game in
Go history and the margin is still disputed. -
Three days, three buildings. Day one went to move 89 at a private house in Naniwa. Play
continued somewhere else up to move 141, and finished at an inn at Nakanoshima. Move 127
was played at the third place, so the famous ears turned red in a rented room. -
The engine’s alternative is a corner move. C18. Everybody remembers a grand central
play and the machine wants to take points in the top left. -
K11 is one line above tengen. The most famous move in Go history misses the exact centre
of the board by one line. -
Genan never got the Meijin title. He died in 1859.
-
Shusaku’s Castle Game record was 19 wins, 0 losses, from 1849 to 1861. Thirteen years
without a loss. He died of cholera in 1862 at 33, after staying in the Honinbo house to
take care of the sick.
What I take from this
The engine does not make move 127 holy and it does not condemn it. It costs about a point,
which is what normal strong moves cost.
What the numbers show is that this was the best game of Shusaku’s life by a large margin.
What the story shows is that a man who could not read a board could read the game from
somebody’s ears.
Neither of those needed permission from a machine.
