Consistency between /standard and /standard/mainline endpoints

[Pawel Kacprzak]

I'm using the API for the Reddit bot and people just reported it provided a wrong continuation line so I'm trying to examine it further.

Here is the FEN: 7k/8/1R6/R7/8/8/8/4K3 w - - 0 1

Now, http://tablebase.lichess.ovh/standard?fen=7k/8/1R6/R7/8/8/8/4K3_w_-_-_0_1 says that the best move is Ra7 and indeed it is, but mainline returned by http://tablebase.lichess.ovh/standard/mainline?fen=7k/8/1R6/R7/8/8/8/4K3_w_-_-_0_1 starts with Rg5 which is obviously not the best move (still a win but not the quickest one).

I'm basing the bot's internal logic on the assumption that if the /standard endpoint returns dtm different than null then /standard/mainline returns mainline leading to mate with the given dtm. Now I'm wondering if this assumption is even valid?

[Niklas Fiekas]

Syzygy tablebases guarantee winning won positions, but not the shortest path to mate (only the shortest path to the next winning pawn move or capture). To quote https://syzygy-tables.info/:

DTZ optimal play is not always the shortest way to mate (DTM) and can even look unintuitive: For example sometimes pieces can be sacrificed to reset the fifty-move counter as soon as possible. However, unlike DTM it achieves the best possible result even with the fifty-move rule.

Now back to this API.

/standard/mainline provides the pure DTZ mainline. Checkmating on the next move is valued no differently than making a pawn move and checkmating later.

/standard move ordering takes some auxiliary information into account:

https://github.com/niklasf/lila-tablebase/blob/c996c1b6c243ab64a301201eeabfe89d0121d76b/src/main.rs#L344-L352

In particular it orders also by DTM values obtained from 5-piece Gaviota tablebases, if Syzygy tables say that the 50-move rule is not relevant.

So the endpoints are different in that regard, the latter closer to what you're looking for, and I believe neither exactly it.

[Pawel Kacprzak]

Thanks for the explanation. I knew about the DTZ approach but didn't know what exactly the /standard/mainline returns, now it's clear. I think I'd just follow your suggestion to generate a short mainline continuation with the shortest path to mate (if DTM is available):

Alternatively, how about always just doing sth like 5 /standard requests in sequence? It's a little bit of HTTP overhead, but avoids the expensive full mainline probes and works with the existing limits.