Every game review on Chessblood labels each move with a classification: a blue brilliant, a green best, an orange mistake, a red blunder. The labels are useful only if you know what they measure. This article gives the exact rules the engine applies, in order, with the thresholds written out.
Win probability, not centipawns
The first thing to understand is the unit. Classifications are not based on how many centipawns a move lost. They are based on how many percentage points of win probability it lost, from the point of view of the player who made the move.
The reason is that centipawns do not mean the same thing in every position. Consider two moves that each lose exactly one pawn's worth of evaluation:
- From +0.3 to -0.7, the game goes from roughly equal to clearly worse. In win-probability terms that is a drop from 52.8% to 43.6%, about 9 points, an inaccuracy on the edge of a mistake.
- From +8.0 to +7.0, nothing has happened. The mover's win probability goes from 95.0% to 92.9%, about 2 points. The move is still excellent.
Converting the engine score to a probability first makes the labels say what a human means by "mistake": the move changed the likely result. The conversion curve is the one described in the article on what chess accuracy is. The short version: 0.00 is 50%, +1.00 is about 59%, +3.00 is about 75%, and anything beyond +15.00 is capped near 100%.
The thresholds
Chessblood publishes the full rule set, and the same numbers are shown on the accuracy calculator. The rules are checked from top to bottom; the first one that matches decides.
"Decided" means the mover's win probability before the move was above 95% or below 5%. Great and brilliant moves are never awarded in positions that were already won or already lost, because finding the only move when you are a queen up is not an achievement.
The quiet categories
Book moves are neither praised nor punished. The label simply means "this position is in the opening table", which is why a slightly dubious but well-known sideline can still be marked book. Once you leave the table, every move is graded on its merits.
Forced moves are excluded from everything. If you are in check and have one legal reply, the engine does not congratulate you for finding it, and your accuracy figure does not include it either.
Best, excellent and good are the normal moves of a decent game. Best is defined by a threshold, not by matching a single move: in a position with three roughly equal continuations, any of them can be best.
Inaccuracy, mistake and blunder are the three sizes of error. Ten points is roughly the difference between an equal position and one where you are a clean pawn down. Twenty points is the difference between equal and losing.
Missed win describes a specific disappointment: you were winning, you let most of the advantage go, but the game is not lost. "Mistake" undersells it and "blunder" oversells it.
Great: the only good move
FEN
r1bq1b1r/ppp2kpp/2n5/3np3/2B5/5Q2/PPPP1PPP/RNB1K2R b KQ - 1 7The Fried Liver Attack gives a clean example. After 1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 d5 5.exd5 Nxd5 6.Nxf7 Kxf7 7.Qf3+, Black's king has to walk forward with 7...Ke6, defending the knight on d5. Stockfish puts the position at about -1.0 after it, which is 41% for Black. The second-best try, 7...Ke8, allows 8.Bxd5 and drops to roughly -2.8, or 26%. The gap between the best and second-best move is almost 15 points, well above the 10-point threshold, so 7...Ke6 is marked great.
Notice that great does not mean the position is good. Black is worse after 7...Ke6. The label says that every alternative was significantly worse, which is exactly the kind of move that deserves a mark in your notes.
Brilliant: a sound sacrifice
FEN
r1bq1rk1/ppppbppp/2n1p3/3nP3/3P3P/3B1N2/PPP2PP1/RNBQK2R w KQ - 1 7The oldest recorded Greek gift comes from Gioachino Greco's notebooks: 1.e4 e6 2.d4 Nf6 3.Bd3 Nc6 4.Nf3 Be7 5.h4 O-O 6.e5 Nd5 7.Bxh7+ Kxh7 8.Ng5+ Bxg5 9.hxg5+ Kg8 10.Qh5 f5 11.g6, and mate follows on h8. Run through the brilliant checklist:
- Is it the top move or within 0.5 points? Stockfish rates 7.Bxh7+ at about +4.9, level with the quiet 7.c4. Yes.
- Does it sacrifice at least two pawns' worth? A bishop for a pawn. Yes.
- Was the position undecided? White's win probability before the move was about 86%, below the 95% cutoff. Yes.
- Does the mover keep at least 35% afterwards? Comfortably.
- Was the mover already five pawns up? No, material was level.
So 7.Bxh7+ is brilliant. The example also shows something people often get wrong: brilliant does not mean "the only winning move". It means a real sacrifice that the engine confirms is sound in a position that was still a game. A speculative sacrifice that happens to work against a weak reply is not brilliant, because the engine measures the move against best defense.
Blunder: from worse to lost
FEN
rn1qkbnr/ppp2p1p/3p2p1/4N3/2B1P1b1/2N5/PPPP1PPP/R1BQK2R b KQkq - 0 5After 1.e4 e5 2.Nf3 d6 3.Bc4 Bg4 4.Nc3 g6 5.Nxe5, Black is already in trouble. The best move, 5...dxe5 6.Qxg4, leaves Black a pawn down at about -2.2, or 31% win probability. Taking the queen with 5...Bxd1 is answered by 6.Bxf7+ Ke7 7.Nd5 mate, so Black's win probability goes from 31% to 0: a drop of 31 points, a blunder by a wide margin. A blunder is measured from where you were, not from equality: Black went from worse to mated, and that counts.
Why different sites give different counts
If you run the same game through two review tools, the counts will not match, and that is not a bug. Several things differ:
- Thresholds. Where one tool draws the line between inaccuracy and mistake is a design choice. Chessblood's lines are listed above.
- Depth and engine version. A move that is best at depth 14 can be third at depth 22. Deeper analysis usually produces fewer "best" labels and more small inaccuracies, because the engine finds more precise moves to compare you against.
- The book definition. A larger opening database marks more moves as book, which removes them from grading entirely.
- Brilliant criteria. This is the category with the most variation between tools, because "sacrifice" has to be defined in code: how much material, confirmed over how many moves, with what safety margin.
Compare your own games across time using one tool at one depth, and the numbers become meaningful.
How to use the labels
Do not chase brilliants. They are rare by design and mostly appear in games where you were already doing well. The numbers that predict your results are the count of inaccuracies and mistakes in positions that were equal, because those are the moments where the game was still there to be won. Open your game review, filter for the orange and red moves made before the position was decided, and spend your study time there.