Absurdly long forced mate evaluations in some games

[dan9er]

I've been processing some database dumps (namely to find a forced-mate to win% curve a la lichess-org/lila#11148), but I found some games have evaluations with absurdly large forced mate evaluations. Like, into the millions or even billions of full-moves. (For comparison, the longest known forced mate is #549 ignoring the 50-move rule.)

Here are some games from December 2025 my preprocessing script has found since I added a check for this:

• https://lichess.org/m8ktnmt5#121
• https://lichess.org/8ehpTJZs#100
• https://lichess.org/bz61a9mQ#71
• https://lichess.org/iVUEQC5Z#99
• https://lichess.org/yo5dkSjL#79
• https://lichess.org/5mu39F0E#90
• https://lichess.org/QbzNk4uo#112
• https://lichess.org/51cSsamg#100
• https://lichess.org/myUJjgY6#99
• https://lichess.org/z0sldOqC#108
• https://lichess.org/OLgYrSS0#90
• https://lichess.org/WfheBnwg#100
• https://lichess.org/nTE2OipQ#113
• https://lichess.org/nTE2OipQ#115
• https://lichess.org/johafN5y#120
• https://lichess.org/c50AmF8W#86
• https://lichess.org/JNrWQ3xZ#104

It's still processing right now, around half done. (Done now, remaining games added to the above list.) I also use the Gaviota DTM tablebases for <=5 piece endgames without castle rights before checking the %eval comment, so there may be more.

Here are also some points in my processed datasets from before I added the check. I don't have the game IDs, and I can't guarantee these are all from separate games.

• November 2025
  • #16066, White won, Blitz
  • #-1019672388, Black won, Blitz
• October 2025
  • #1049214217, White won, Rapid
  • #1002858761, White won, Blitz
  • #844849361, White won, Blitz
  • #1021005695, White won, Blitz
  • #-1015295096, Black won, Rapid
  • #1020408300, White won, Blitz
  • #-1072213380, Black won, Bullet
  • #-806759701, Black won, Blitz
  • #981038985, White won, Blitz
  • #997796652, White won, Bullet
• June 2022
  • Clean, as far as I know.

[dan9er]

I did a bisect search on the database dumps and it appears the first game to have a messed up mate evaluation is https://lichess.org/dJd2k6q2#102 from 2025-09-02.

The last commit before this is 187e165, which bumps dependencies and the Stockfish version. I suspect this is either a Stockfish bug or a bug in how some Rust library is parsing its output.

[niklasf]

Thanks for reporting and for the analysis -- that's very helpful to narrow it down. All the sample games you collected come from one particular client. I've contacted the contributor to see if we can debug the issue.

[dan9er]

Hello niklasf, any updates on this?

If you want I can make a PR to add a known issue notice to https://github.com/lichess-org/database.

[niklasf]

I did not find the root cause, but lichess-org/lila@60e7165 forced an update above the version that appears to have produced it.

So if the rest of January doesn't have further occurences, just documenting the issue should be fine.