You were winning. The eval bar said +3.2, then +2.8, then you glanced at the clock: thirty seconds. Eight moves later your flag fell in a position the engine still called winning. So what does an engine evaluation mean, if it said nothing about what was actually going to happen?
It never promised to. A Stockfish evaluation is not “White is a pawn and a half better.” It answers a different question: how does this game end if both sides play perfectly from here? Only an engine plays perfectly. You don’t, and neither does your opponent.
+1 is not a pawn, it is fifty percent
The hundredths used to be called centipawns, and +1.00 honestly meant “roughly an extra pawn.” Since version 16, Stockfish has tied its scale to win probability. Per the engine’s own docs, +1.00 means a 50% chance of winning in Stockfish-versus-Stockfish play, and 0.00 does not mean “equal chances,” it means an almost certain draw.
First consequence: +0.5 is nothing. It is a position where White is slightly more comfortable but a careful Black holds. Half of all openings are evaluated that way, and nobody calls them bad for Black.
Second consequence: the scale is non-linear. The gap between 0.0 and +1.0 is the gap between “draw” and “half the games are won.” The gap between +5.0 and +6.0 is nothing: both positions are won, the engine is just refining by how much.
The same number means different things in different phases
An extra pawn in a king-and-pawn endgame is almost always a win, and the engine will show +1.5 or more. The same extra pawn in a middlegame with queens on, an open king and loose pieces is also +1.5. But the second position is twenty moves away from a win, each of them a chance to ruin everything, while the first is ten moves of technique you either know or don’t.
So always read the number together with the material on the board. The fewer the pieces, the closer the eval is to the result. The more pieces, the more it means “who is more comfortable” rather than “who wins.”
And a third variable the evaluation does not contain at all: the clock. +3 with thirty seconds against a human who has five minutes is zero. The engine needs a millisecond to find the right move; you need exactly the seconds you don’t have. How we measure real playing strength without that distortion is a separate story, but a board in time trouble has nothing to do with the eval bar.
Translating into human
Keep a four-line table nearby, it covers nearly everything you will see in a game review:
- −0.5 to +0.5 — equal. Play the position, not the number.
- around +1.5 — something to fight for. Not a win, but the right to a long piece of work, described in the article on converting an advantage.
- +3 and above — won with technique. The only way to lose is to hurry or overlook counterplay.
- “mate in 6” — the most dangerous line of all. It means a mate exists, not that you see it. You still have to calculate, and calculate precisely: in a mating net one inaccurate check turns +M6 back into +2.
Notice what the table does not contain: the difference between +1.7 and +2.1. For a human it does not exist. If you are arguing with yourself about which of two moves is better and the eval difference is under half a pawn, the engine does not know the answer for you. It knows the answer for itself.
Where the engine hurts you
Picture two positions. The first is +0.3: White is slightly better, but to keep that plus you must find three only-moves in a row, each against intuition. The second is 0.0: a calm symmetrical structure where any sensible move holds the balance.
The engine prefers the first. A 1400-rated human goes wrong on move two of the first position and lands at −1.5, while in the second they play forty more moves without risk. The better position for you is the one you know how to play, not the one with the bigger number.
Hence the most common distortion in post-game analysis: a player looks at the graph, sees “it was +2, now it’s +0.8,” and concludes they blundered. Sometimes yes. Sometimes they deliberately traded queens and went from a +2 position that demanded precision into a +0.8 endgame with an extra pawn that wins on technique. The engine calls that a loss. A coach calls it the right decision.
How to read the graph after a game
Don’t look for the peak. Look for two kinds of points: where the line crashed, that is your blunder, and where it sat flat at +1.5 for a long stretch while you never moved it, that is your gap in planning. The first is fixed with attention, the second with knowledge, and those are different kinds of training.
Everything else, the half-pawn wobbles, is noise. The engine sees thirty moves ahead and penalises you for missing a move that none of your opponents would have found either. Beating yourself up over it has no point.
That is why the game review on 99Chess counts not pawns but the drop in win probability after each move: inaccuracy, mistake and blunder are defined by how many percentage points you gave away, not by how many hundredths the number shifted. And the coach explains in words why a move was bad, because the number alone never will.
An engine evaluation answers the question of how a game between two perfect players ends. You are the one at the board. Your question is different: which of these positions can you actually carry to the end?