An Evaluation Number Needs a Sentence Beside It
A numeric score is a compressed output of a search/evaluation system. It does not tell you by itself whether the advantage comes from material, king safety, activity, a forced tactical sequence, a long-term structural feature, or several interacting factors. Good analysis translates the number back into chess language.
Centipawn-Style Scores Are Normalized Evaluations
The familiar +0.50, +1.00 or -2.00 notation is best treated as a scale for engine preference rather than literal material accounting. Modern Stockfish, for example, normalizes its displayed centipawn-style evaluation using a WDL model; its documentation describes +1.00 in terms of a calibrated self-play winning probability, not “White owns exactly one extra pawn.” That calibration belongs to the engine/model context and should not be generalized into a universal human probability.
Two positions with the same numerical score can be strategically very different. One may be a stable extra pawn with low risk; another may be a temporary initiative requiring several accurate moves.
Perspective Must Be Explicit
A sign only has meaning if you know the convention used by the interface. Many GUIs present scores from White's perspective; other contexts may use side-to-move or another convention. When recording analysis, store the perspective rather than relying on memory.
Mate Scores Are a Different Semantic Category
A mate score says that the current search has a mating result represented by the engine/interface convention. Treat it separately from centipawn arithmetic. Do not describe mate as merely “+100.” If the mate claim matters, inspect the line and verify that the search is stable enough for the purpose.
WDL Is Model Output, Not Your Personal Win Probability
Win/Draw/Loss output can be more intuitive than a single score because it exposes the engine model's outcome distribution. But it is still calibrated from engine data and assumptions. It does not directly answer, “What are my chances of beating this particular opponent with twelve minutes left?” Practical difficulty, player strength, time control and familiarity live outside that model.
Version and Configuration Matter
Engine architecture, networks, normalization and options evolve. Durable analysis records the engine/version when a precise claim depends on it and avoids turning version-specific UI behavior into timeless chess doctrine.
Compare Scores Only Within a Known Context
Evaluation scales are not guaranteed to be identical across engines, versions or interfaces. Even when two tools use centipawn-like notation, their calibration and display conventions can differ. For precise longitudinal research, keep the engine/version consistent or treat the scores qualitatively rather than pretending they share a universal ruler.
The Shape of the Position Matters More Than Decimal Precision
A stable +0.8 position with a clear technical plan and a volatile +0.8 tactical position can impose completely different human tasks. Always pair the output with a description of the position's demands: forcingness, king safety, conversion difficulty, available counterplay and sensitivity to error.
Reference Framework
An engine is a chess-search/evaluation program. It is not the GUI that displays it, not the cloud service that hosts it, not a game database, and not an endgame tablebase.
Engine outputs must be interpreted as model/search outputs with a perspective and configuration.
A number from an engine is evidence produced by a tool—not a complete chess explanation.
What a Chess Engine Is
A chess engine is software that searches chess positions and evaluates candidate continuations to choose/analyze moves.
Engine vs GUI
The engine performs chess search/evaluation; the GUI sends commands, manages positions/settings and presents output. A GUI can host different engines.
Engine vs Cloud Service
A cloud analysis service is an application/infrastructure layer that may run one or more engines remotely with service-defined resources/settings.
Engine vs Database
A game database stores/retrieves games, positions and metadata. It does not become an engine because it displays statistics or stored evaluations.
Engine vs Tablebase
A tablebase stores exact low-material endgame information within covered conditions. It is conceptually different from heuristic engine search.
UCI Concept
UCI is a protocol/interface for communication between chess engines and GUIs/controllers. It is not the engine itself.
Centipawn-Style Evaluation
Centipawn-style evaluation is a normalized numeric score used to express engine preference/advantage on a pawn-like scale. Its semantics are engine-specific and model-based.
Why +1.00 Is Not Literally One Pawn
+1.00 is not literally one pawn of material. Modern Stockfish normalization ties the score to its WDL/self-play calibration rather than a simple material count.
Stockfish Normalized Evaluation
Current Stockfish documentation states that its normalized +1.00 corresponds to about a 50% self-play win probability in its calibration context. This is version/model context, not a universal chess law.
Evaluation Perspective
Every numeric evaluation needs a perspective convention. Store whether the sign is shown from White's perspective, side-to-move perspective or a GUI-defined convention.
White Perspective vs Side-to-Move
White-perspective GUIs keep positive as White-favoring; other engine/protocol displays can be interpreted differently. Never infer perspective from the sign without knowing the interface convention.
Mate Scores
Mate scores indicate a mating result under the current search/proof representation, commonly with a distance convention. They should be interpreted separately from centipawn scores.
Mate Is Not Centipawn Arithmetic
Mate is not a very large centipawn advantage. Once mate is established, the semantic result is qualitatively different from heuristic CP arithmetic.
WDL
WDL output presents model-estimated win/draw/loss proportions/categories associated with the engine's evaluation model.
UCI_ShowWDL
Stockfish exposes WDL analysis through the UCI_ShowWDL option.
WDL Is Engine-Model Output
Engine WDL is engine-model output, calibrated from engine self-play/search context.
WDL Is Not Human Win Probability
Engine WDL is not a direct probability that a particular human player will win/draw/lose the position. Human strength, time control and practical difficulty are different variables.
Evaluation-Output Practice
For evaluation-output practice, compare CP, mate and WDL examples and state the perspective, model limits and what each output can and cannot prove.