looking for suggestions where to look

[flok]

Hi,

I rewrote part of the threading of my chess-program.

To see if anything changed, I took 87 FENs and calculated the best-move for each with a depth of 10.
I then verified that the number of nodes visited is exactly the same of before-the-change and after-the-change in single-threaded mode.
Also the time spent doing this verification has a difference of less than 1%.
Furthermore, I also let both versions play a game against themselves and compared everything there's to compare (fens, scores, etc): all the same.

Yet if I run a match between the old and the new version (in cutechess-cli or fastchess), then I get a negative difference of 39 elo with 3025 games played! (new version is weaker than the old version)
What on earth could cause this difference?

[Henk]

Maybe go back to the previous version and redo a small step.

[flok]

Pondering. The older version cheated: if it wasn't explicitly told via UCI commands to not use ponder, it would start doing so after the first "go ...".

[Brunetti]

Pondering. The older version cheated: if it wasn't explicitly told via UCI commands to not use ponder, it would start doing so after the first "go ...".

Are you talking about Dog? Which version?

Alex

[flok]

Pondering. The older version cheated: if it wasn't explicitly told via UCI commands to not use ponder, it would start doing so after the first "go ...".

Are you talking about Dog? Which version?

Alex

Dog yes, all published versions.