Chess analysis often reports centipawn loss: the difference between the engine's evaluation and the evaluation after your move. It is intuitive, but it treats all pawns as equal, which the position does not.
The problem with centipawns
Going from +0.20 to -0.80 is a 100-centipawn loss and a serious problem. Going from +8.00 to +7.00 is also a 100-centipawn loss but changes nothing practical. A linear centipawn scale overstates errors in already-decided positions and understates errors near equality.
Win probability
Win probability converts an evaluation into an expected score between 0 and 1. A move that drops the expected score from 0.55 to 0.45 is a real swing; a move that drops it from 0.99 to 0.98 is not. ChessUrge uses this model to classify blunders, mistakes and inaccuracies in full-game reviews.
| Evaluation | Rough win probability |
|---|---|
| 0.00 | 50% |
| +0.50 | 55–60% |
| +1.50 | 70–75% |
| +3.00 | 85–90% |
The exact curve depends on the model, but the shape is consistent: near equality, small evaluation changes matter a lot; at the extremes, they matter less.
Use both numbers
Centipawns are useful for comparing quiet positions and understanding engine preferences. Win probability is better for judging practical mistakes and explaining why a move was a blunder. A good review shows both: the evaluation and the expected-score swing.
Open a full review in the library to see the move labels and the evaluation behind them.
See the whole game, not just the blunder.
Choose a review depth, see the maximum cost, and get move-by-move accuracy with blunder, mistake and inaccuracy markers.