When Two Engines Give Different Answers

Set the same position in front of two strong engines and you can get two different numbers, occasionally two different best moves. Neither is broken. The disagreement has a cause, and the cause tells you which answer to prefer.

Reason one: they searched different amounts

The most common explanation, and the least interesting. One engine ran to depth 22 and the other to depth 30, or one had a second of thinking time and the other had a minute.

Check the depth before assuming anything else. If they differ, give both the same budget and see whether the disagreement survives. About half the time it does not.

Reason two: the position is genuinely unresolved

Some positions do not have a settled value at reachable depths. Sharp middlegames with several plausible sacrifices, or endgames whose result depends on a long race, will produce different numbers from any two searches that took different routes.

The signature is instability within each engine, not just between them. Run one engine twice at different depths. If its own answer moves substantially, you are not looking at a disagreement between two engines; you are looking at a position neither has solved.

Reason three: they are built differently

When these two disagree on a quiet, closed position where the plan runs twenty moves deep, the network-heavy engine is usually the one to believe. When they disagree in a position full of checks and captures, the search-heavy engine usually is. That is a rule of thumb rather than a law, but it fails less often than picking whichever number you preferred.

Reason four: one of them is looking at a different position

Worth checking before the interesting explanations. A FEN copied with the wrong side to move, or with castling rights stripped, produces a legal position that is not the one you meant. Two engines given two different positions will of course disagree.

This is more common than it sounds, particularly when a FEN has been through a screenshot, an app and a clipboard.

What to do about it

For a position you care about, three steps settle nearly every case.

1. Give both the same depth and the same time. Much disagreement is an artefact of unequal budgets.
2. Look at whether each engine's own answer is stable as depth increases. An unstable answer is not an opinion, it is an unfinished search.
3. If the disagreement survives both, look at the lines rather than the numbers. Two engines that disagree about the evaluation but agree about the moves are not really disagreeing.

And if the difference is under half a pawn, stop. Two strong engines differing by 0.4 in a complicated position is normal variation, not a puzzle to solve.

A case worth recognising

One disagreement pattern comes up often enough to name. A search-heavy engine reports a position as roughly equal; a network-heavy one reports a clear advantage for the side with a long-term structural edge, such as a protected passed pawn in a closed position.

Both are describing something true. The first is saying that no forcing line converts the advantage within its horizon. The second is saying that across many similar positions, this structure wins. Over the board, against a human, the second assessment is usually the more useful one, because nobody defends a passive position accurately for forty moves.

The reverse case is a sharp position with three pieces hanging. There, a network's judgement about which side stands better is worth much less than a deep search's verification of what actually happens, because the answer is concrete and calculable rather than a matter of tendency.

Do not average them

The instinct when two numbers differ is to split the difference. That produces a figure neither engine would endorse and which describes no position.

If the two agree on the best move and differ on the number, the number was never the point. If they disagree on the move, one of them is going to be shown wrong by deeper analysis, and the way to find out which is to search deeper rather than to compromise.