Is Stockfish reliable as far as draw conclusions is concerned?

Question

I just stumbled across this endgame : https://www.365chess.com/game.php?back=1&gid=134885&m=24

Basically, Black has only a Queen, White has Queen and Rook. Black resigns.

But here comes the tricky part:

If I put the game in a computer (here Stockfish 15), it does not find the winning ending game for White, even after a depth 70+ analysis. (I used stockfish-mac , https://github.com/daylen/stockfish-mac/wiki/User-Guide/). Black maintains perpetual check.

Though I'm not an expert of chess endgames, it does not seem to me that this position is documented as a theoretical draw. And this was an actual resignation from chess Grand Master and International Master. I suppose they knew what they were doing, even if Stockfish nor I can't find the winning position.

So basically I assume the game was losing for Black but Stockfish can't prove it.

I don't know what to do with this information.

a) Is it a Stockfish limitation? Should the program be completed by an endgame database, as seems to be hinted on the software help page (see syzygy) ?

b) But then, does this mean that without an endgame database, Stockfish 15 is faulty ? These bases are TeraBytes, which makes them pretty unusable for most people

c) One might argue that I've come across a very specific case, but is it really ? If Stockfish can't handle this basic situation, can it be trusted for other cases ? After all, many game evaluations rely ultimately on endgame evaluation.

d) By the way, the 50 move draw rule which is a 100 depth Stockfish analysis seems out-of-reach for Stockfish in reasonable time on a normal computer

e) Maybe Stockfish is right, this game was a draw, but then why did the chess masters decide it was not, and why is this not a theoretical draw documented (I couldn't find anything on google after a short search)

f) If Stockfish is right, that means that 365chess.com includes deceiving results, which is disappointing to me as I use it to check after strength of openings

Any expert insight about this would be appreciated.

Thanks

Answer 1

These bases are TeraBytes, which makes them pretty unusable for most people

I suspect you don't completely understand how these tablebases are used.

According to the Syzygy webpage, a total of 939 MB of data is required for 5-piece tablebases, 156 GB for 6-piece and 16.7 TB for 7-piece.

So, first of all, you could have a 5-piece tablebase for less than 1GB and your engine would be able to look up KQR vs KQ (5 pieces) in the tablebase and get an immediate answer. Not prohibitive at all in terms of data space required. With a decent sized SSD the 156 GB for a 6-piece tablebase is also reasonable.

However that's not the end of the story. If you have a 6-piece tablebase that will still turbocharge your engine when analysing endgames with 7 or more pieces because whenever the engine analysis reaches a position with 6 pieces it will be able to immediately look it up in the tablebase with no further calculation required.

So basically I assume the game was losing for Black but Stockfish can't prove it.

The position you give is a draw in the tablebases so your assumption here was wrong.

a) Is it a Stockfish limitation? Should the program be completed by an endgame database, as seems to be hinted on the software help page (see syzygy) ?

It is a limitation, yes, in the same way that an engine without an opening book is limited. In both cases the engine will take longer (and may never reach) to calculate the best moves.

b) But then, does this mean that without an endgame database, Stockfish 15 is faulty ?

No. It just means that you will get better results from it if you use tablebases. As already pointed out, smaller 6-piece or even 5-piece tablebases will still improve performance.

Answer 2

Browsing through other answers here I found this website https://www.shredderchess.com/online/endgame-database.html

based on a 1TB 6-pieces endgames database, that gives the answer:

The game is a draw.

Black should not have resigned. Sorry for the noise.