Fortresses, and Why Engines Miss Them

There are positions where one side is a rook ahead and the game is a dead draw. An engine will report those at plus five or worse, and no amount of ordinary searching corrects it.

This is the clearest case where you should trust your own eyes over the number.

The structural reason

A fortress is defined by what cannot happen. To prove one, you would need to establish that across every sequence of moves, of any length, no progress is available.

A search cannot do that. It examines lines to some depth and scores the positions it reaches. Every position it reaches still has the extra rook, so every position scores well. The engine finds nothing wrong because nothing goes wrong within its horizon, and the draw only becomes visible at a horizon it never reaches.

Deepening the search does not fix this. At depth 40 the engine has verified that the advantage persists for forty plies, which is exactly what a fortress guarantees.

The ones worth knowing

The wrong-coloured bishop case is the one every player meets. A rook pawn on the a or h file with a bishop that does not control the promotion square is a draw, whatever else is on the board, provided the defending king reaches the corner. Engines handle this one correctly now, because it is common enough to be recognised specifically, which tells you something about how the rest are handled.

How to spot one during analysis

Three signs together are close to conclusive.

The evaluation is stable across depth but the position does not change. Play the engine's own moves for twenty plies. If the resulting position looks materially identical to the starting one, with the same blockade intact, the advantage is not being converted because it cannot be.

No pawn breaks exist. Look for any move that changes the pawn structure. If every pawn is fixed and no capture opens a file, there is usually no way in.

The defending king cannot be driven from its post. If the key defensive square is covered and cannot be attacked more times than it is defended, the position is closed.

What to do about it

Tablebases settle the question exactly for small numbers of pieces, and if the position is down to seven pieces or fewer the answer is not an estimate at all. Above that, you are on your own.

The practical move is to stop asking the engine and start asking whether a plan exists. Write down what you would actually do to make progress. If the answer is a plan, test it. If the answer is that you would shuffle and hope, the engine's number is describing material rather than the position, and it is time to take the draw.

The mirror image

The same blindness works against you. If you are the defending side, an engine will tell you that you are losing by five pawns in a position you can hold indefinitely, and that is discouraging in a way that costs games.

Players resign fortress positions. They also decline to steer into them, because the evaluation on the screen says the position is hopeless when in fact it is the only place the game is saveable.

If you are much worse and can see a structure where everything is fixed and your king covers the entry square, that is worth heading for regardless of what the number says. The number is measuring what you have, not whether it can be used.

Recognising the shape early

Fortresses do not appear by accident. They are reached by a defender who spots the possibility a dozen moves ahead and arranges for it.

The ingredients are consistent: a fixed pawn chain that the attacker cannot break, a defending king on a square that covers the only entry point, and at least one defensive piece that cannot be driven away or exchanged off. When you are defending badly, look for whether those three can be assembled. They frequently can, and the engine will not suggest it.