Puzzle where engine contradicts solution

Hi all, I just ran across a puzzle where the engine analysis suggests a different move than the puzzle solution (lichess.org/training/yJkqw). Is this because the puzzles were generated with a different version of the engine (neural network version)?

In the puzzle, after 17. Nxh4 Qxh4, there are two possible captures for white to win a piece: Bxg4 or Nxh5. The in-browser stockfish prefers Bxg4, but the puzzle solution is Nxh5 - just curious what is the explanation, and if anyone has any insight as to why to prefer one capture over the other? I don't think it's about engine depth - I let it run for a while (depth 37).

I think Bxg4 *should* be the solution instead of what the puzzle has as the solution. IMO it's better to lose the h-pawn than the central e-pawn.

github.com/ornicar/lila/issues/8116#issuecomment-774775621

Mentioned it on github. Thanks for reporting

In this puzzle

lichess.org/training/skrBi

the engine solution only works, because 0-0-0 is forbiden (the king moved in the game). how should the person solving the puzzle know, that 0-0-0 is forbidden? Better remove this puzzle.

but anyway - you are doing great work

thospie

#4 You have the notation on the right.

@Deadban:
thats right. Is it the idea of a puzzle, that i have to check 29 moves of a game if a possible solution is a legal move? In my opinion ...

...and do you have the notation on your mobil app?