Eval: chess engine evaluation
Eval
Definition
Eval (short for “evaluation”) is the numeric score a chess engine assigns to a position to indicate who is better and by how much. By convention, positive numbers favor White, negative numbers favor Black. The unit is typically the centipawn (abbreviated CP), where +1.00 ≈ a one-pawn advantage for White. When a forced checkmate is detected, engines switch from a centipawn score to a mate score, displayed as “M#” (for example, M3 means mate in three moves).
Eval is central to engine-assisted analysis. You’ll see it in post-game reports, opening preparation, and live broadcasts as the moving “eval bar.” Related terms include Engine eval, Engine, Computer move, Best move, and Tablebase.
How engine eval is expressed
Units and signs
- Centipawns (cp): +0.20 means a small positional edge for White; −1.50 means Black is about a pawn and a half better.
- Mate score: “M5” indicates a forced mate in 5 for the side to move; “−M4” indicates Black mates in 4.
- Depth and MultiPV: Engines also show search depth (e.g., depth 30) and multiple principal variations (MultiPV) with competing lines and evals.
Typical scale of advantages
- ≈ 0.00: Dynamic equality or a Dead draw with best play.
- +0.20 to +0.70: Small, “healthy” edge; often a space or structural plus.
- +1.00 to +2.00: Clear advantage; often convertible with accurate play.
- ≥ +3.00: Technically winning for strong players, barring practical counterplay.
Important caveat: Numbers are context-dependent. Fortresses, Opposite bishops endgames, or massive complications can make a “+2.00” practically hard to win, while a “0.00” may be a razor-thin Drawing line that’s not easy to find OTB.
Usage in chess
How players and analysts use eval
- Opening prep: Choosing lines where the engine shows stable equality or a small edge; rejecting “Dubious” moves with persistent negative evals.
- Post-game analysis: Spotting turning points (eval “swings”) after a Blunder or finding a hidden Zwischenzug.
- Practical decisions: Preferring a slightly inferior eval if it offers better Practical chances, a common human strategy versus clinical engine lines.
- Endgames: Trusting tablebase-backed 0.00 or win/loss verdicts; understanding when the engine’s “+0.00” actually means a forced draw.
Live and online
The “eval bar” popularized on broadcasts and platforms gives a fast visual of who’s better. Beware: reacting to every fluctuation can be misleading in sharp, tactical positions where the “truth” stabilizes only at higher depth.
Strategic and historical significance
From handcrafted evaluations to neural networks
Early engines relied on handcrafted evaluation terms (king safety, piece activity, pawn structure). Landmark systems like Deep Blue (Kasparov vs. Deep Blue, 1997) used massive computation plus expert-tuned eval heuristics. Modern engines combine powerful search with learned evaluation: AlphaZero (2017) and Leela Chess Zero use neural networks producing value estimates, while Stockfish’s NNUE (from 2020) blends fast search with a neural net inside a classical framework. Result: more accurate, “human-like” evals, especially in messy middlegames.
Impact on theory and practice
- Opening theory: Many “refuted” or “dodgy” lines have been rehabilitated (or buried) by stable engine evals at depth. Entire repertoires hinge on whether evals settle near 0.00 or give a persistent edge.
- Style and training: Players learn to balance engine-approved accuracy with human feasibility—knowing when to follow a “Computer move” and when to choose simpler paths.
- Endgame truth: Endgame tablebase verdicts refine engine evals to perfect outcomes in 7-piece (and beyond) settings.
Examples
1) Small, stable edge out of the opening
In the Queen’s Gambit Declined, White often retains a slight pull that engines quantify around +0.20 to +0.40.
Sample line (illustrative eval ≈ +0.25 for White):
2) Tactical blunder and an eval swing to mate
Scholar’s Mate shows how a single mistake can send eval from roughly equal to a forced mate:
Line: 1. e4 e5 2. Qh5 Nc6 3. Bc4 Nf6?? 4. Qxf7#
Before 3...Nf6??, eval is near equality; after the blunder, engines display “M1” for White. Patzer sees a check… and finds mate!
3) “0.00” that still requires precision
Some endgames are theoretically drawn (eval 0.00) but demand exact moves—classic recipe for a Swindle if the defender slips. Think rook-and-pawn vs. rook endings where a single tempo loses the Lucena position race.
Reading eval like a pro
Practical tips
- Use eval to compare candidate moves, not to “play the number.” If a +0.60 line is nightmarish to navigate, a quieter +0.20 with clearer plans can be better OTB.
- Check MultiPV: If several moves hold equality, don’t waste time chasing a tiny eval edge that’s fragile at higher depth.
- Watch depth: Early, shallow evals can flip after deeper search. “Best move” claims are provisional until the horizon is pushed back.
- Respect resources: Fortress ideas, perpetual checks, or stalemate traps can neutralize large centipawn leads. See also Fortress and Perpetual.
- Endgames: When available, consult tablebase lines—these override heuristic evals and show exact results.
Common misconceptions
- “0.00 means boring.” Not necessarily—sometimes it’s a razor’s edge draw requiring only moves.
- “+2.00 is easy.” Not if the defender has counterplay or if the position is a technical grind.
- “Engines always find humanly playable wins.” Engines may prefer narrow, computer-only lines; consider Practical chances.
Interesting facts and anecdotes
- Eval bar drama: The “rollercoaster” bar has become a broadcast star—volatile in sharp lines like the Najdorf or King’s Indian.
- Kasparov vs. Deep Blue, 1997: A milestone where machine evaluation and calculation shocked the chess world and reshaped preparation forever.
- Neural revolution: AlphaZero’s policy/value approach reframed evaluation as learned intuition; Stockfish’s NNUE fused that intuition with classical search speed.
- “M” moments: Spectators love sudden flips to “M#,” often following a quiet Intermezzo or a stunning Queen sac.
Related and see also
Quick reference
Eval thresholds (rule of thumb)
- ≈ 0.00: Equal with best play
- +0.20 to +0.70: Small edge
- +1.00 to +2.00: Clear advantage
- ≥ +3.00: Winning with correct technique
- M#: Forced checkmate in # moves
Use eval as a guide, not a goal. The best practical move is often the one that balances objective strength with human clarity.