Plus 1.20 Explained: Reading The Eval Bar Without Panicking

The engine says +1.20. You are playing White, you have a bishop pair and a slightly loose king, and the bar on the side of the board has slid comfortably into your half. So you have a winning position, right?

No. You have a position where, if both sides played the next thirty moves perfectly, White would end up a bit better than a pawn ahead in material-equivalent terms. That is a completely different claim, and the gap between those two readings is where most club players’ engine habits fall apart.

This is a post about calibration: what the numbers actually denominate, what each band means for someone rated 1400 rather than 3400, and how to stop letting a red bar talk you into resigning a drawable rook ending.

The number is a forecast of a perfect game, not a probability about yours

Stockfish’s evaluation is expressed in centipawns. One hundred centipawns is nominally one pawn, and Stockfish’s UCI output gives you cp 120 for what your GUI renders as +1.20. That much most people know.

What gets skipped is the second half of the definition. The evaluation is the value of the position at the end of the principal variation, after both sides have played the engine’s best moves to the search horizon. It is not a measurement of the position in front of you. It is a measurement of a position twenty-odd plies away that neither of you will ever reach, discounted back and printed as a single number.

Try this in the Lichess analysis board. Load any middlegame, open the engine, and hover the evaluation. Lichess shows you the principal variation alongside the number. At depth 24 you might see:

+1.20  24/35  Nxe5 Bxe5 Rxd8+ Rxd8 Qc2 Bxh2+ Kh1 Be5 Qxc6 bxc6

That +1.20 belongs to the position after ...bxc6, ten moves down the line. If your opponent deviates on move two of that sequence, the number is already stale. If you deviate, it was never about you at all.

Newer Stockfish builds muddy this further by defaulting to a normalised scale where +1.00 is pinned to roughly a 50% win expectation for the engine itself at that node. Stockfish 16 onwards, UCI_ShowWDL gives you the honest version directly:

info depth 26 score cp 120 wdl 331 601 68

Three numbers per thousand: 33.1% win, 60.1% draw, 6.8% loss. That is Stockfish’s estimate of its own result against itself. A +1.20 is a position where a superhuman engine wins one game in three and draws most of the rest. Between two 1400s, the same position is decided by whoever blunders first, and someone will, usually within eight moves.

Recalibrating the bands for a human

Here is the translation table I actually use when reviewing games with students. The left column is the engine. The right column is what it means in a game between two humans who each get one bad idea per twenty moves.

Notice how much wider the practically-undecided zone is than the engine’s. Stockfish considers +0.80 a real achievement. For you it is a hint about where to look, not a lead to protect.

The band that causes the most damage is the third one, roughly +0.80 to +1.50, because it is where the bar looks decisive on screen and isn’t. A pawn-and-a-bit does not win games below master level. Conversion does. Go and count how many of your own games were at plus one at move twenty and finished as losses. If you have played 200 rated games, the honest number is probably fifteen or twenty.

Worked example one: the rook ending you were about to resign

Position: you are a pawn down in a rook ending, four pawns against five, all on the kingside, rooks on the board, kings active. Stockfish says -1.10.

The reflex is to see a minus, feel behind, and start playing passively while you wait to lose. That reflex is wrong, and here is how to prove it to yourself instead of taking my word for it.

Open the position in the Lichess analysis board and check the tablebase panel. If you are down to seven pieces or fewer, Lichess queries the Syzygy tablebases directly and will tell you Draw or Win in 41 with no evaluation number at all, because the answer is known rather than estimated. Rook endings are famously draw-heavy: a single extra kingside pawn with rooks on is drawn far more often than not. The engine’s -1.10 is not a forecast that you lose. It is a statement that a perfect opponent extracts a pawn’s worth of pressure from you, and a perfect opponent is not sitting across the board.

Second check: switch Stockfish to MultiPV 3 in the Lichess engine settings (the gear icon, then set the number of lines to 3). If the top three moves read -1.10, -1.15 and -1.18, the position is a plateau. You can play almost anything reasonable and stay in the game. If they read -1.10, -3.40 and -3.60, you are balanced on one move and need to find it. Those are wildly different practical situations that the single-number display collapses into the same -1.10.

That MultiPV spread is the single most useful habit in this post. One number tells you the height of the cliff. Three numbers tell you how wide the ledge is.

Worked example two: plus two that evaporates

Take a Sicilian Dragon Yugoslav Attack sideline where Black is up a piece for two pawns, engine says -2.10 for White. White is objectively worse by the engine’s standard, and at depth 30 it stays worse.

Now look at what has to happen. Black’s king sits on g8 behind a hacked-open h-file. White’s practical plan is Rdg1, g4-g5, h4-h5, and sacrifice something on h5 or h6. Black defends by finding a precise sequence of five or six only-moves. Drop the engine’s depth to 12 in the Chess.com analysis board, or just watch the number at low depth, and you will see something like +1.40 for White, before deeper search finds Black’s resources and flips it.

Shallow depth is not a bug here. It is a rough model of a human who has not calculated to the end. If the evaluation swings hard between depth 12 and depth 30, the position is tactically sharp and full of chances for both players regardless of what the final number says. A position that reads -2.10 at every depth from 10 to 35 is genuinely lost. A position that reads +1.40 at depth 12 and -2.10 at depth 30 is a fight.

This depth-sensitivity business is where a lot of the real skill in engine use lives, and it deserves more room than I can give it here. The pillar piece on reading engine output goes through depth, node counts and the specific ways engine output misleads you at length.

What to actually do with the number

Three rules, and then a training plan.

Rule one: never read an eval without reading the principal variation. The number alone is a rumour. The PV is the evidence. If you cannot follow the PV and understand why each move is played, the eval has taught you nothing except a feeling.

Rule two: check the spread before you judge the position. MultiPV 3, always. A position with one good move is a position you will probably get wrong. A position with five moves within 0.30 is one where your general understanding is enough.

Rule three: treat any eval under ±1.50 as a fight. Not a lead, not a deficit. A fight. Below 1900 the result correlates far more strongly with who blunders next than with who was slightly better at move twenty-five.

For the training plan, the useful version is uncomfortable. Take twenty of your own losses. Find the move where the eval moved by more than 1.50 in one ply, which Lichess flags as a blunder with a ?? and Chess.com tags in red. Those are your real errors. Now look at what happened for the ten moves before each blunder: the eval will usually have been drifting in the 0.00 to -0.60 range while you made harmless-looking moves. That drift is the actual lesson, and it is invisible if you only chase the big red bars.

Then take twenty of your wins where you were at +1.00 or better by move twenty. Count how many times the eval dropped back under +0.50 before you won anyway. If it happens in more than half of them, you have a conversion problem, and conversion is trainable: play the winning positions out against Stockfish limited to 2000 Elo (Lichess offers exactly this through its Stockfish levels, and Chess.com’s bots cover the same range) and see whether you can hold the advantage for fifteen moves.

The eval bar is a very good instrument pointed at a question you did not ask. It answers “what would happen between two perfect players.” You need “what is likely to happen between me and this specific opponent, both of us tired, both of us with eight minutes left.” Nothing in the bar answers that, but the PV, the spread across the top three moves, and the behaviour of the number across depths get you most of the way there.

Start with MultiPV 3 on your next review session. It costs one click and it will change what you notice.