Exact TT scores and alpha/beta

[hgm]

That sounds like it would be better to search multiple times to start with. Just as an inefficient way to prune less...

[Evert]

What else is there, other than the search window and the depth? How wide the window is?

Node type. Stockfish at least makes a distinction between PV and non-PV nodes, but it also has a "score is improving" thingy that I'm not sure of how it works, but it makes the reduction decision based on more than just the current node.

EDIT: also, lots of history-like stuff that affects move ordering. Shouldn't affect the result in a stable search, but Stockfish' search isn't stable.

[Michel]

also, lots of history-like stuff that affects move ordering. Shouldn't affect the result in a stable search, but Stockfish' search isn't stable.

One cannot fail to be amazed with the amount of elo SF has gotten out of its many history mechanisms. Recently "Capture history" was introduced to order capture moves. That made little sense to me as folklore knowledge dictates that captures are already quite easy to order well. Yet here we are...

[hgm]

Node type. Stockfish at least makes a distinction between PV and non-PV nodes,

That should not matter. You can only get a hash cutoff in a PV node from an exact score, which must also be produced by a PV node.

but it also has a "score is improving" thingy that I'm not sure of how it works, but it makes the reduction decision based on more than just the current node.

EDIT: also, lots of history-like stuff that affects move ordering. Shouldn't affect the result in a stable search, but Stockfish' search isn't stable.

This all sounds very fishy. If you lose strength by not searching again what you already searched before when you accidentally get there again trhough a transposition... Then is should gain you even more strength when you would always search that node twice, first time you get there.