BUG - Zero Inaccuracies evaluation, despite blunder

Sometimes the Computer Analysis shows my game to have 0 Inaccuracies, 0 Mistakes, and 0 Blunders. I feel thrilled with my "perfect" play, until I look closer and see that's not the full story.

The following recent game is an excellent example of this. The game itself doesn't matter too much (my stronger opponent was mimicking queen odds in a series of fun casual games), however notice the evaluation from my move 23 to 24.



I very obviously blundered my queen on move 24, and the local eval drops my advantage by 10. However, the server Computer Analysis still shows 0 Inaccuracies, 0 Mistakes, 0 Blunders.

My catastrophic failure totally slipped through the cracks, so this must be a bug, and a critical one at that.

Nope, not a bug, and nowhere near critical. Look at the evaluation graph.

Not a blunder. If they take your queen you have mate in 2. 24. Rxf2 Re1+ 25. Rf1 Rxf1#

Maybe you lost internet connection right when the local eval was yet only one ply deep? :D

@ChesswurmOTB That's exactly what I was going for when I played 23..Qf2, but it doesn't work after 24. Bd2, which guards e1.

@Toadofsky The severity of it aside, can you help clarify why you think it's not a bug?

The graph shows going from -17.3 to -7.7; a drop of more than a queen's worth of significance. This should logically be at least an "inaccuracy" or a "mistake". In the past, I've seen the engine characterize a missed opportunity for +0.5 advantage as a "mistake". Here we see the analysis ignore a ten point drop like nothing ever happened, which is what I'm terming a bug.

Also you helped highlight an unrelated minor bug in how the evaluation graph is shown. I'll document that separately, but basically the vertical axis is not evenly/uniformly distributed and is visually inconsistent with the evaluation. The drop from move 23 to 24 should not be a small change, it should visually be a major drop (over half the vertical size).

#5 Again, this is a matter of opinion... you're not the first to ask, and so I'll explain yet again, based upon github.com/ornicar/lila/issues/4705

I could continue asking questions, such as "What does -17.3 mean?" and "What does -7.7 mean?" and most players couldn't provide a strict definition. You played a move in a completely winning position which gave away 10 points but resulted in another completely winning position. That's not a blunder, although it's not the engine's preferred move. In fact, engines such as Stockfish, Shredder, Rybka, Leela, etc. may provide different evaluations for your moves, but likely every engine would agree that Black is completely winning.

In general, non-mate scores are incomplete since each position can have only one class (regardless of distance to mate):
* Winning for White
* Drawn
* Winning for Black

I continue to recommend to Lichess that engine evaluations should be replaced with winning odds % values, but I'm the only person in favor of the idea so I've mostly given up on the suggestion. Here is a game by the last person who claimed to understand chess engine evaluations. I'm not saying he's wrong, but he still has something to prove:
www.chessgames.com/perl/chessgame?gid=1891857

@Toadofsky I think the last blog entry mentioned winning chance as a new evaluation feature. forgot if it was about being planned or implemented...

Thanks @Toadofsky, you have an interesting idea in general.

This bug I present fundamentally calls into question the reliability and trustworthiness of the computer analysis. If dropping a queen for no reason doesn't change a 0/0/0 tally, then I can't imagine how many other aspects of players' games are being carelessly ignored.

@Talmanian Thanks for the feedback, although even blindfolded I could win both of those positions, so this isn't a bug.

"I very obviously blundered my queen on move 24"

If you actually analyze the position after best play by both sides 25. Rxf2 Rxf2 26. Be3 Re2 27. Bg1 Bxf4 28. c4 Nd6 29. cxb5 axb5 30. Nh5 Be5 31. Ng3 Rxb2 32. Re1 Re8 33. Rd1 Nc4 34. Ne4 d6 35. Nf2 Rg8 36. Rd3 Rf8 you'll find Black's attack completely crushing. It's probably forced mate both before and after you *traded* for a rook plus space plus initiative. If I do end up filing a bug, it'll be that, "Computer analysis does not solve for mate in every position."

Hi, I think in this case the way things currently are are probably best (don't think -17 to -7 should be considered inaccuracy since both of them are completely winning)