Stockfish says "blunder", but why? Turning evaluation into training

Short answer

The inaccuracy, mistake and blunder labels come from an arithmetic operation: the centipawn drop between the evaluation before and after your move, compared against fixed thresholds. That makes the labels useful for triage and insufficient for learning, because the same drop can mean completely different things depending on the position — and it never tells you the human cause of the error.

The essentials in a few lines

• A centipawn is one hundredth of a pawn; the evaluation is a prediction of the result with perfect play, not a grade for your move.
• Typical thresholds are around 50, 100 and 200 centipawns for inaccuracy, mistake and blunder — and they vary between platforms.
• The same 300-centipawn drop matters far more in a balanced position than in one already won or already lost.
• Converting to win probability explains why a mistake at +8 barely changes anything and a mistake at 0.0 changes everything.
• The label tells you the size of the damage. The cause — didn't see it, miscalculated, misjudged, ran out of time — is always yours to determine.

What the evaluation actually measures

The choice of the pawn as the unit is historical and practical. Internally the engine adds up dozens of factors — material, king safety, pawn structure, mobility, weak squares — and needs to express the result as something interpretable. A centipawn is one hundredth of that unit, and it is the granularity at which platforms do their arithmetic before rounding to the number you see.

Three practical consequences follow. First: the evaluation assumes perfect play from both sides, which rarely describes your next game. Second: it is always relative to search depth — the same position can be +0.4 at one level and 0.0 ten plies deeper. Third: 'mate in 5' is not a numeric evaluation but a verdict; once the engine shows that, the pawn scale no longer applies.

It is also worth clearing up a common confusion: an evaluation of 0.0 does not mean 'a dull game' or 'nothing to do'. It means that with perfect play the expected result is a draw — which is perfectly compatible with a position full of traps where the human side goes wrong four times out of five.

Inaccuracy, mistake and blunder: the ruler behind the labels

Because the thresholds are fixed and the scale is not linear in importance, the same number means different things in different contexts. Losing 300 centipawns from a 0.0 position trades a draw for a likely loss. Losing the same 300 from +8.0 changes nothing: you are still winning trivially. The 'blunder' label appears in both cases.

That is why platforms have also adopted the conversion to win probability. On that scale, what is measured is how much of your expected result the move destroyed — and it matches human intuition better: the mistake that hurts most is the one made near equality, where every tenth of a pawn has real consequences.

One practical implication for anyone reviewing games: counting blunders is a poor progress metric. An improving player usually trades many cheap blunders in already-decided positions for fewer expensive ones in balanced positions — and the raw count can even rise while the rating rises too.

A concrete example in five moves

This is the most economical illustration of the central point: the engine does not evaluate intention, plan or appearance, only the resulting position. 3...Nf6 obeys every opening principle a beginner is taught — it develops, attacks the enemy queen, prepares castling — and still loses on the spot, because no principle applies when there is a concrete mate threat.

The human cause here is always the same and has a name: failing to check the opponent's threat. The player generated candidates from their own goals and never asked what the opponent's last move started threatening. It is a process error, not a knowledge error — and the training that fixes it is a checking routine, not more opening theory.

Notice what an automatic review would deliver here: 'blunder, 3...g6 was best'. Correct, complete and useless as training, because your next game will not have a queen on h5 on move three. What carries over to the next game is the question you failed to ask, and that question is not in the report.

From the label to the human cause

That is the translation the engine does not make, and it decides the value of the whole analysis. The label classifies the damage; the cause classifies you. Two games with the same number of blunders can call for completely different training — tactical patterns in one case, clock management in the other — and nothing in the automatic report distinguishes them.

A useful signal for separating perception from calculation: if, reviewing the position calmly with no engine, you find the correct move in under a minute, the problem is almost certainly neither knowledge nor calculation. It is process under pressure — candidate generation, threat checking or the clock. If you cannot find it even calmly, then it is knowledge, and the training is different.

Record the cause alongside the position, not in a separate list. A position annotated 'never checked the mate threat on f7' is a reusable exercise two weeks from now; a list saying 'I made 4 blunders' is not training material at all.

• Always review the position before the move, never the one after it.
• Question 1: was the right move on my list? If not, the training is perception.
• Question 2: did I reject it by miscalculating, or by misjudging the end of the line?
• Question 3: how much time was left? Below 20%, treat it as a clock error until proven otherwise.
• Write the cause into the PGN as a comment, next to the position.

Turning evaluation into exercises

Five positions with the same cause make a coherent training session: five where you did not check the opponent's threat, or five where you stopped calculating before the opponent's retreating move. Five positions with different causes make a collection, and a collection fixes nothing.

The two-week interval is not arbitrary: it is what separates recognising from remembering. A position reviewed the next day is solved by memory of the episode; the same position two weeks later tests whether the pattern actually stuck. That is the difference between reviewing games and training with them.

Finally, a better metric than counting blunders: track the average evaluation of your position at move twenty across twenty games. It measures the accumulated effect of your opening and early middlegame decisions, and it moves far more steadily than any error count — which depends on how much your opponents let you get away with.

Frequently asked questions