Learn After the game Lesson 3 of 5
What a move costs
Inaccuracies, mistakes and blunders: weigh moves yourself, find the only move, and see the chance nobody took.
No move makes a position better
This sounds wrong, and it's the key to everything an engine says about your moves.
The engine's score for a position already assumes you'll play the best move in it. So when you do, the score doesn't go up: it stays where it was. And when you play anything else, it goes down.
That means a move is never measured by what it gains. It's measured by what it gives away: the win chance before it, less the win chance after it. That's the move's cost.
A move that costs nothing is as good as any there was. A move that costs 30 turned a position worth 70% into one worth 40%.
Names for the sizes
By tradition, costly moves get names and marks, and each site draws the lines where it likes. These are the ones the reports here use, in points of win chance:
| It costs | It's called | Mark |
|---|---|---|
| Under 5 | A good move | |
| 5 or more | An inaccuracy | ?! |
| 10 or more | A mistake | ? |
| 20 or more | A blunder | ?? |
Under five is noise: the engine's own number moves that much as it thinks. From ten up, something real happened.
Try it: weigh some moves
On this board, every move you can play has already been weighed by Stockfish. Play one, and you'll see the win chance it leaves, what that costs, and the reply that makes it cost that much, drawn in red.
Then take it back and try another. The moves you've tried are listed under the board with what each one keeps.
Notice how the good moves bunch together at the top, a point or two apart, and then the floor falls away.
The same slip, a different price
What a slip costs depends on when you make it. Take losing a pawn for nothing.
| You were | You are now | It cost |
|---|---|---|
| +1.0, at 59% | 0.0, at 50% | 9 |
| +8.0, at 95% | +7.0, at 93% | 2 |
The same pawn, and one is nearly a mistake while the other is nothing at all. When you're a queen up, almost any move still wins. When the game is level, small things decide it.
So the moves worth your attention are the costly ones played while the game was still open. A piece dropped in a position that was already lost changes nothing, however bad it looks.
Try it: when only one move will do
Usually there are several good moves, and choosing among them hardly matters. Sometimes there's exactly one, and everything else loses.
These are the moments that decide games, and nothing on the board announces them. An engine spots them by comparing its best move with its second best: when the gap is wide, the position is critical.
Here's one from King and pawn endings, in numbers.
Try it: the chance you didn't take
A move can cost a great deal without losing anything at all.
When your opponent makes a mistake, your win chance jumps at once, before you've done anything. If you then play an ordinary move and don't take what was offered, the chance drops back, and the engine counts the drop against you.
These are the hardest mistakes to find by yourself, because nothing bad ever happens on the board. The game just goes on.
What a cost doesn't tell you
- It's measured against perfect play. A move that costs 25 against Stockfish may cost nothing against the people you play, if the punishment is one nobody at your level finds. And a move that costs 8 may set you a problem you get wrong every time.
- It doesn't say what kind of mistake it was. A missed fork and a slow plan can cost the same.
- One costly move is an accident. It tells you about that game. It takes many games to tell you about yourself.
A report here goes after all three. It asks a second neural network, trained on games between people, what each mistake really costs against opponents at your level. It sorts your mistakes by kind. And it reads enough of your games to find the ones that keep happening. You can see an example report to look at how that comes out.
The last lesson of this level is how to go over a game of your own.