What the Stockfish evaluation number actually means

The number is a pawn count that stopped being a pawn count

A Stockfish evaluation of +1.24 means: at the depth it searched, White's position is worth about one and a quarter pawns more than Black's. Positive is always White, negative is always Black, regardless of who is to move. Centipawns are hundredths of a pawn, so +124 and +1.24 are the same claim in different units.

The catch is that modern Stockfish does not add up pawns. Since version 12 the evaluation comes from a neural network, so +1.24 can appear in a position where material is dead level — it is the net's summary of king safety, structure, activity and whatever else it has learned, expressed in pawn-sized units because that is a scale people can read.

Stockfish's own documentation says the output is normalised so that +1.00 corresponds to roughly a 50% chance of winning in its own play at typical test conditions. That is a statement about Stockfish beating Stockfish. It is not a statement about you beating the person you are paired with at 10+0.

Mate scores are a separate scale. M5 means forced mate in five moves, -M3 means you are getting mated in three. These are not centipawns and cannot be compared to them.

Misreading 1: treating the eval as a win probability for you

A -0.9 does not mean you are losing. It means that with perfect play from both sides the position is slightly worse for you, and that perfect play is being defined by an engine that will never get tired on move 38.

Between 800 and 1800, positions swing several pawns in both directions within a few moves. An evaluation of -0.9 in a sharp middlegame is, in practice, closer to a coin flip than the number suggests. The useful threshold is not 0.0, it is the point where the position stops having play in it.

The opposite error costs more. A +0.35 out of the opening is not an advantage you can bank. It is the engine saying the position is basically balanced and White is marginally more comfortable. If you are replaying your openings to squeeze +0.35 into +0.5, you are optimising the part of the game that is not deciding your results.

Misreading 2: assuming the eval drop names your mistake

The graph falls off a cliff at move 24, so you look at move 24. Often the move that lost the game was move 19, and move 24 was the first position where the engine could see the consequence.

Engines mark the move at which the evaluation changed, not the move that caused it. A drop from +0.4 to -2.1 sometimes means you hung a knight. Sometimes it means you traded the wrong pair of pieces three moves earlier and only now is there no way to defend d5.

When you see a large drop, step back four moves and ask what changed structurally in that window. If you cannot find anything, it probably was a one-move tactic, and that is a different problem with a different fix.

The same mistake runs the other way. The engine flagging your move as a mistake does not mean the alternative was findable. If the refutation is a seven-move line ending in a quiet retreat, filing it under "I blundered" teaches you nothing.

Misreading 3: reading the number without the line or the depth

A single evaluation with no principal variation attached is almost useless. The number tells you the position is +2.5. The line tells you why, and the why is the only part you can carry into your next game.

Depth matters more than most people check. An evaluation at depth 15 and the same position at depth 30 can disagree by several pawns in tactical positions, and the shallow one is not a rough version of the deep one — it can be flatly wrong. If a browser or phone gives you a number in half a second, that is a shallow number.

Two known weaknesses are worth carrying around. Engines still misjudge fortresses, sometimes showing a two-pawn advantage in a position no human or engine can break. And a 0.00 does not always mean a balanced position — it can mean the engine has found a forced repetition inside an otherwise wild fight.

Turning on multiple lines fixes a lot of this. If the top three moves all evaluate within 0.2 of each other, the engine is telling you the choice is yours. If the best move is +1.8 and the second is -0.4, there is exactly one idea in the position and you need to know what it is.

Misreading 4: treating all lost centipawns as the same problem

Average centipawn loss reads like a single dial you can turn down. It is not. It is a mean, and it sits on a distribution with a long tail.

Across 189,777 engine-classified moves from 6,059 games analysed on Mated, the average move gives away 94 centipawns against the engine's best — but the median is far lower, because a small number of very bad moves carry most of the damage. Your ACPL is largely a report on your worst four or five moves of the game.

The same data set shows 3.8% of moves played are outright blunders and another 4.2% are mistakes. That is roughly one move in twelve being a real error, before you count the 18.4% classified as inaccuracies. Those inaccuracies are noise in the sense that they rarely decide anything on their own. The one move in twelve is what you actually play for.

So when you compare your 62 ACPL to someone else's 45, you are mostly comparing how many pieces each of you hung, laundered through an average. Look at the count of moves losing more than 150 centipawns instead. It is a cruder statistic and it tells you more.

Where the damage actually happens

Errors are not spread evenly through a game. In that same set of 189,777 engine-classified moves across 6,059 games, 9.1% of middlegame moves are a mistake or worse, against 2.3% in the endgame.

Two readings of that are both true. The middlegame has more pieces, more candidate moves and more ways to be wrong. And games that reach an endgame are, by selection, games where nobody collapsed earlier — so endgames in this data are calmer than endgames in general.

What it means for your analysis time is simple. If you are spending your review reading opening evaluations and endgame technique, you are working on the cheap parts. The move where +0.8 became -3.0 is almost always somewhere between move 15 and move 30, in a position with no forcing move and no obvious plan.

This is the part that is tedious to do by hand, which is what Mated is for — it runs Stockfish over every position of your last games in the browser, finds the drops, and rebuilds them as positions you have to solve. But you can do the same thing manually with a board and an hour. The method matters more than the tool.

The question to ask at every drop

When you find a move that lost more than 150 centipawns, the engine has done its job and you have to do yours. Ask one question: did you see the opponent's reply at all, or did you see it and judge it wrongly?

These are different failures with different fixes. "I did not see it" means you are not checking your opponent's checks, captures and threats before you commit — a process problem, fixable with a habit. "I saw Nxd5 and thought I was fine after Bxd5" is a judgement problem, and it is fixed by understanding the position type, not by looking harder.

Most players never separate the two. They log every eval drop as a blunder, feel bad, and resolve to concentrate more. Concentration fixes the first category and does nothing for the second.

Nobody has measured well how the split changes as you improve — it would need data on what players were actually thinking, not just what they played. My own guess, and it is only a guess, is that vision failures dominate below 1200 and judgement failures take over somewhere around 1600. Sort your own drops into the two piles for ten games and you will know which one is yours.