if The AI see x stones (maybe 3 or 5) in a raw change negatively our chance to win, it should tell us the starting point of the bad exchange.
In my experience, it does exactly that. There is no defined “exchange” in go. Sometimes the start and end of an exchange are clear, sometimes it’s just a rough term. Sometimes even exchanges are interrupted by forcing moves the opponent throws in. So you can’t really say “this exchange started here”. But the AI tells you exactly the point in the game where the score estimate shifted this or that way. Which means if there was an ongoing exchange, it went fine up until that point.
i look many analyse of your AI and i found it not always the case .
I was saying many time myself : “ok i shouldnt play here at all”
the at all mean they are a first stone i put in a area i should not put.
Or i should play better before the moves AI show on top 3 “+10% moves”.
The algorithm to find the free Top 3 moves has indeed the problem of being a few moves late in some cases. This is known, but difficult to address without increasing the load on the server.
If the AI gives your move a high negative win rate change, this is the move where you got wrong. If the difference to the suggested move is minimal, the bad play happened probably a few moves earlier.
The full review for supporter usually finds the point where the you got wrong.