a possible wrong 9x9 book line

[hhfhhfhfh1]

i just researched this line in 9x9 book and i supposed it to be a 6 depth wrong line, so i thought it worth an issue

from this situation, the white move is forced by black according to the book until this position

using the recent network s1150 ,i analyzed this position and it gives a different winrate

unfortunately, the book of this position didn't expand this move , so i use b18 for 9x9 to analyze this position and conformed my conclusion,

the b18 net ignore the trick here and gives a 50% winrate until 100k visits

[hhfhhfhfh1]

addation:this line also seems to be wrong ,though im not that sure than the first one.
the second one is the same situation but chooses the forth choice instead of the third

i use net s1150 and come to this situation

the white move was pretty forced by black though not that forcing than first one
and the 9x9 book seemed to ignore this move

i use the b18 for 9x9 net and it did not found this move with more than 10m visits

and i continue the line and it found it with more than 500k visits

[lightvector]

This is excellent, thank you! Please keep posting more variations with major errors if you find them, I can use them to expand the book and this will likely correct the evaluations.

[hhfhhfhfh1]

i checked the 8x8 book and found a wrong(?) line and some fake scores.
this line is analyzed and proved to be wrong, however the black score has hint that it might be a wrong line ,so i don't think it worth a shocking discovery.

the first choice here was wrong, the s1150 suggests the second choice is the only way to win. Besides this position, there are other positions has two acceptable moves, but they only matters the order and don't affect the result, so i omit them.

the analyze of s1150, which position 9x9-18b net ignore the trick here

the same position in 9x9 book for compare.

the following positions' winrates are correct, but the score was wrong. due to they choose the "others"'s score which has only 7k visits instead of the first move. it can be easily fixed by adding 20k-30k visits, or let the book ignore the "others"'s score and winrate when there are more than 2 moves. there are more examples, but i don't want my comment to be too long ,so i just post two of them.

[hhfhhfhfh1]

A very very strange position

after very long time analyze, B6, G7 turned out to be wrong,G5 is so unlikely to be correct that i didn't want to spend time to follow,if it turned out to be a draw i would be very surprised

however, i tried so many visits and nets(2~10m per net and >10 nets),they initially showed me that it's a white winning position.

at first, when I used the 60b network, its evaluation suggested a drawn position.
I initially attributed this to miscalculation - which was indeed the case. But when comparing its errors with the latest 28b network, I found the 28b network also produced miscalculations. The specific problematic positions have been overwritten in my records, and I can't recall the exact scenario. Ultimately, when running millions of visits from the position shown above, subsequent variations would be led astray, maintaining a 99% win rate until suddenly dropping to 50% at certain blind spots. This issue was significantly mitigated when starting analysis directly from later positions.
The core problem was KataGo's mistaken assumption that White's upper group required Black to spend an extra move to capture, resulting in a 1-2 point loss for Black. In reality, Black has a tactical sequence to simultaneously prevent White's living formation while securing both side hane (kake) threats, ultimately leading to a drawn game.
so i think it's a very interesting positon, where initially analyze showed draw, when adding more calcultions, it turned to be white winning,but after some deep and artificial analyze, it became draw again.
i paste the sgf below and it also includes some other positions which the book may miscalculate.but i haven't finish it yet.

8x8-1.txt

[hhfhhfhfh1]

update:this position turned into a white winning position with latest net analyze, i'll show it later in my sgf which includes the book's miscalculate