Why does stockfish make these strange recommendations?

Question

The FEN is:

r2qk1r1/ppp2p1p/2n1bn1Q/4p1B1/8/2P4P/PP1NpKP1/R5NR b kq - 0 1

Clearly black is winning because the white bishop can be captured (with a discovery available if white recaptures). However, the stockfish eval seems strange - why would it suggest that playing an intermediate Rg6 is better? The continuation is Qh4 Rxg5 Qxg5 anyway, which is just a slower way to do the same thing. Wouldn't it be better to play the most direct move Rxg5?

Then I am also confused about the 3rd recommendation of e1=B+, why would it suggest to underpromote there?

Answer 1

I think that in the case of the immediate ...Rxg5, Stockfish has seen one move deeper than in the version with ...Rg6 first. As a result the evaluation is slightly different - 0.1 less, in this case. So it ranks ...Rg6 first. It doesn't really matter either way, it's just an artifact of how engines work.

Same for ...e1=B+. It gets captured immediately, so it's effectively the same move as ...e1=Q+ that would also be captured immediately. Which one it happens to show is essentially random, it doesn't have a builtin preference for promoting to a queen in cases where it doesn't matter at all.

Answer 2

No particular reason. Note the scores, -8.9 and -8.8, practically indistinguishable. Stockfish simply saw the same sequence of tactical response with and without the move ...Rg6. The line ...Rg6 came first due to random fluctuation.