A centipawn is one hundredth of a pawn's value — the unit engines use to evaluate positions. An evaluation of +1.5 means White is better by the equivalent of one and a half pawns; negative numbers favor Black. Move-quality labels like blunder and mistake are defined by how many centipawns a move loses.
Every evaluation bar you have ever watched is speaking in centipawns. The engine reduces everything it understands about a position — material, king safety, piece activity, pawn structure — to a single number, denominated in hundredths of a pawn.
The convention is simple. Positive numbers mean White stands better, negative numbers mean Black does. +0.3 is a nudge; +1.5 says White is effectively a pawn and a half up, even if the pawn count on the board is equal; +5 usually means a whole piece. Around ±0.5 and below, engines are describing a position that is, for human purposes, balanced.
When a forced checkmate appears on the horizon, engines abandon centipawns entirely and display M5 or #-3: mate in five for White, mate in three for Black. There is no meaningful exchange rate between +9 and mate — one is an assessment, the other is an announcement.
It's also worth knowing that modern engines internally think in terms of winning probability and convert to centipawns mostly for our benefit. That is why the same +2 advantage feels different in a sharp middlegame than in a simplified endgame — and why analysis tools increasingly weigh eval changes by their effect on the expected result.
The practical reason to care: every move you play is graded by centipawn loss — the gap between your move and the engine's best. Small losses are inaccuracies, larger ones are mistakes, and game-changing ones are blunders. Your average centipawn loss over a game (ACPL) is a decent one-number summary of how accurately you played.
But averages hide what matters. A game with 20 ACPL can contain one catastrophic blunder and 39 perfect moves. Improvement comes from finding the handful of big-loss moves and understanding them — which is what game analysis is actually for.
That is the principle behind Chessdock: it uses engine evaluations of your games to locate the exact moments your evaluation collapsed, and rebuilds them as training puzzles.