How to study chess openings with an engine without becoming a slave to the evaluation bar

Short answer

Studying an opening with an engine needs a different procedure from analysing a finished game: what matters is not the best move but what produced the number. This guide defines the evaluation-bar trap, gives the matrix that turns evaluation bands into positional concepts, shows two cases measured in Stockfish 18 where the naive reading fails — a capture that loses seventy hundredths between depth 20 and depth 30, and a position rated 0.00 that ended in mate sixteen moves later — and closes with a forty-minute protocol for studying a new line.

The essentials in a few lines

• The evaluation-bar trap is mistaking evaluation for difficulty. The engine measures what the position is worth with perfect play from both sides; it does not measure how many only-moves you will have to find to hold it.
• Depth changes the size of the advantage, not just the decimal. In the Traxler Defence, 5.Nxf7 reads +0.78 at depth 20 and +0.08 at depth 30, while 5.Bxf7+ reads +1.24 and +1.20 at the same two: at the shallower setting both captures look playable, at the deeper one only one of them is worth anything.
• A level evaluation is not a level game. Move 17 of Fischer–Myagmarsuren reads 0.00 and the game ends in mate on move 33: the number measured the position, not the gap between the player with a plan and the player without one.
• A difference below two tenths of a pawn in an opening position is no reason to change repertoire. In the sharp position measured for this guide, changing only the engine settings — threads, memory and number of lines — moved the number by six tenths at the same depth 20.
• When studying a line, use multi-PV and compare the top two or three moves against each other. The gap between them says more than the first one's absolute value: two moves within a tenth means a choice of style; one move alone by half a pawn means an only-move worth memorising.

The evaluation-bar trap

The confusion is understandable, because the bar is the most visible part of any chess interface and the only one that answers in a second. The problem starts when it becomes the criterion for choosing a repertoire: someone abandons an opening because the engine gives the opponent +0.25 and adopts one that gives +0.10, without noticing that they traded a position they understand for one they have never played. Fifteen hundredths of a pawn do not exist in the practice of a player whose games are decided by other things.

There is a second effect, quieter and more expensive. Studying an opening by reading only the engine's first line teaches sequences rather than structures: you memorise that the move is 9.Bf4 without knowing why, and on move 12, when your opponent leaves the script, the memorised information does not help. It is the same failure the guide on analysing games with Stockfish describes in post-game work, moved into opening preparation.

The answer is not to switch the engine off — it is the most reliable tool in existence for checking a line. The answer is to change the question you ask it. Instead of «what is the best move?», three questions the engine answers just as readily and that produce transferable knowledge: what happens to the number when this piece leaves that square, how far apart are the top two moves, and what happens to the evaluation when I increase the depth.

Turning the number into a concept

The technique for finding the owner of the number is simple and rarely used: change the position and measure again. Take the suspect out — move the isolated pawn to a defensible square, restore castling rights, close the diagonal of the bishop pair — and ask for the evaluation once more. If the number falls to near zero, you found the cause; if it does not, your hypothesis was wrong and the advantage lives somewhere else. Two minutes per hypothesis, and the output is a concept rather than a move.

The table below is the rough map that guides that search. It is not an exact conversion rule — the same evaluation means different things in different positions, which is exactly why the right-hand column describes a procedure rather than a verdict. What it fixes is the commonest error: reacting to a two-decimal number as if it were a scoreline.

One order of magnitude is worth fixing in mind, because it saves time. In opening positions, differences below two tenths of a pawn are inside the noise: moving from depth 20 to depth 28 usually shifts the number by more than that. Switching lines over 0.15 is optimising something you cannot measure; switching over 0.60 is a different conversation, and even then the useful question is how many only-moves defending the worse side requires.

Depth, and settings, change the number

The case measured for this guide is the Traxler Defence, which arises after 1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 Bc5. At depth 20, Stockfish 18 rates the capture 5.Nxf7 at +0.78 and 5.Bxf7+ at +1.24 — two moves that look about equally promising, with a modest preference for the second. Stop the analysis at that setting and you conclude that the Traxler is in trouble and that choosing between the two captures is a matter of taste.

At depth 30 the gap between them is of a different order. The knight capture falls to +0.08, essentially nothing, because the engine now sees 5...Bxf2+!, the counterblow the variation is famous for; the bishop capture holds at +1.20. What looked like a choice between two good moves becomes one move that gives an advantage and one that gives equality — seventy hundredths of difference that the shallower search was hiding.

There is a second lesson, found by accident while checking the first, and it is more useful than the original: the same position after 5.Nxf7, at the same depth 20, reads +0.78 on one thread with 16 MB and a single line; +0.92 on four threads with 512 MB; and +1.41 on four threads with two lines in multi-PV. Six tenths of spread with neither the position nor the depth changing. An evaluation without its settings stated beside it is not a measurement — it is an impression, and it does not compare with anyone else's.

Level evaluation, decided game

The example measured here is Fischer–Myagmarsuren, Sousse 1967, the model game for the King's Indian Attack. At move 17, with the centre closed and both attacks under way on opposite wings, Stockfish 18 reads 0.00: perfectly level. The game ends in mate on move 33, without Fischer's opponent making any single mistake the engine would label decisive — the evaluation slides through accumulated inaccuracies, each of them small.

This pattern has a practical name: asymmetry of difficulty. The attacking side chooses between three good moves; the defending side has to find the one move that does not lose. Both produce the same evaluation and they do not produce the same result — and that difference is exactly what opening study should be hunting for, because it is the part that converts into points.

The consequence for repertoire building runs against instinct: prefer the +0.05 line where you hold the plan to the +0.25 line where your opponent does. The bigger number belongs to whoever knows what to do with it, and the engine does not report who that is. The King's Indian Attack guide works through this specific case, with all three base positions measured at the same depth.

Stockfish, Lc0 and the value of disagreement

In opening study the difference shows up in two places. Positional sacrifices with no immediate return — a pawn given for lasting activity — are usually handled better by the network, which recognises the pattern without needing to see the material come back inside the search horizon. In sharp tactical positions with long forced sequences, wide search is the more reliable of the two, which is precisely why the Traxler case above is a depth problem.

The most profitable use of the two is not to pick a winner: it is to treat disagreement as a marker. Where both engines agree, the position is what the number says and there is nothing left to study. Where they differ by more than half a pawn, you have found a position whose nature is debatable — and those are the ones worth understanding, because they are the ones that produce games decided by understanding rather than preparation.

One honest caveat before building that routine: every measurement in this guide was taken with Stockfish 18 and is stated with the depth used, because comparing numbers across engines demands care that is rarely taken. Evaluation scales are not identical between engines, or even between versions of the same engine, and a «+0.45 from Lc0» is not comparable to a «+0.45 from Stockfish» without context. Compare verdicts and move order; distrust comparisons of decimals.

• Engines agree: position settled, move on.
• Divergence above half a pawn: a position of debatable nature, worth human study.
• Divergence that vanishes with more depth: it was a search limit, not a difference of judgement.
• Never compare decimals across engines or versions without saying which is which.
• Positional sacrifices without immediate return are where networks do best; long forcing tactics are where search does best.

A forty-minute protocol

The first stage looks like wasted time and is what makes the rest work. Play the line out on a board and write, in plain prose, where the pieces go, what pawn structure results and what your plan would be. The point is to record your understanding before it is contaminated: without that record you have no way of knowing afterwards what you learned, because everything the engine shows will look obvious in hindsight.

At stage two, switch the engine on with multi-PV set to three lines and play the whole line through once. Do not write down evaluations: write down only the points where your expectation and the engine's diverge, and the distance between its first and second choice. Two moves within a tenth of each other means there is a choice of style there and you may keep the one you understand. One move alone by half a pawn means an only-move — that one is worth memorising.

The third stage is the hunt for the cause described in the second section: change the position, measure again, find the element. The fourth is the one that turns all of it into something usable during a game: three sentences at most, on the plan, the typical break and the mistake to avoid. A repertoire file with three hundred lines and no sentences is an archive; with thirty lines and ninety sentences it is knowledge — and it is what survives when your opponent leaves the script on move 12.

Frequently asked questions

Openings mentioned in this guide

• Italian Game (Giuoco Piano) (C50)
• King's Indian Attack (A07)
• French Defense (C00)