Defining what is ‘local’ automatically is very hard, and kind of goes back to why AIs were so bad at go for forever: evaluating a position is hard to describe in an algorithm.
In any case I disagree with AI sensei’s approach in putting that much focus on the AI’s top choice. It’s the only thing it can do, not having a live connection to an engine, but I personally think it destroys much of the openness that makes go so amazing. There usually are many good moves.
However, this is an interesting site by KataGo’s author which does something in the space of auto generated problems: https://neuralnetgoproblems.com
Essentially: Find situations where a 14k would make a certain mistake, but a 10k would play correctly, and present these as 14k problems.