Game The short answer
Across 4 independent positions, taking the offered pawn was top of the candidates in 2 of 4. Where it was not, it trailed by at most 0.74 (averaging 0.21). So this is not a move that is simply good or simply bad — the answer changes with the position, which is what the run actually showed.
(“Top” here means top of the candidates compared, not the best move among all legal moves — see how we tested it.)
What we are testing
Take back the pawn you just lost — nothing could be more natural. Yet chess has positions where recapturing is exactly what breaks your position. The branch below, known by the alarming name of the Fried Liver, is the classic example. We compared Black’s candidate moves.
Black is to move, so the numbers are from Black’s point of view. They are negative because White holds the initiative here; what matters is the difference between the moves.
How we tested it
Testing a rule of thumb in one position cannot tell you whether you are seeing the rule or just that position. So we picked 4 positions of different character — open and closed, White to move and Black to move. In each one, every candidate is played and the resulting position is searched. All the numbers are real output, not estimates.
What is compared is a set of candidate moves relevant to the article, not every legal move: the move under test, and the moves a player would normally consider there. So the top row of a table is not "the best move in the position" but the highest-scoring move among those compared.
On top of that, each position is also searched once with no candidate list, and the move the engine itself picks is recorded (it is shown with each position below). This is there to catch a candidate list that quietly decides the answer: the generator aborts if that free choice beats the top candidate by 0.50 or more. Across the current data the largest such difference is 0.18.
Position by position
| Position | Top candidate | Move tested | Gap |
|---|---|---|---|
| The Fried Liver moment (Black to move) | Na5 −0.21 | Nxd5 −0.95 | 0.74 |
| The Queen’s Gambit pawn (Black to move) | e6 −0.32 | dxc4 −0.43 | 0.11 |
| The Evans Gambit (Black to move) | Bxb4 ±0.00 | Bxb4 ±0.00 | same |
| The poisoned pawn (Black to move) | Qxb2 −0.40 | Qxb2 −0.40 | same |
Position 1｜The Fried Liver moment (Black to move)
1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 d5 5.exd5
FENr1bqkb1r/ppp2ppp/2n2n2/3Pp1N1/2B5/8/PPPP1PPP/RNBQK2R b KQkq - 0 5
The highest-scoring candidate was Na5 (−0.21). Nxd5 scored −0.95, a gap of 0.74.
Searched with no candidate list imposed, at the same depth, the engine also chose Na5.
Tap a candidate move to change the board
| Move | Evaluation | Diff vs top candidate |
|---|---|---|
| Na5 | −0.21 | baseline |
| b5 | −0.48 | −0.27 |
| Nd4 | −0.54 | −0.33 |
| Nxd5 | −0.95 | −0.74 |
| h6 | −2.06 | −1.85 |
| Be7 | −3.87 | −3.66 |
Position 2｜The Queen’s Gambit pawn (Black to move)
1.d4 d5 2.c4
FENrnbqkbnr/ppp1pppp/8/3p4/2PP4/8/PP2PPPP/RNBQKBNR b KQkq - 0 2
The highest-scoring candidate was e6 (−0.32). dxc4 scored −0.43, a gap of 0.11.
Searched with no candidate list imposed, at the same depth, the engine also chose e6.
Tap a candidate move to change the board
| Move | Evaluation | Diff vs top candidate |
|---|---|---|
| e6 | −0.32 | baseline |
| c6 | −0.41 | −0.09 |
| dxc4 | −0.43 | −0.11 |
| e5 | −0.54 | −0.22 |
| Nf6 | −0.77 | −0.45 |
Position 3｜The Evans Gambit (Black to move)
1.e4 e5 2.Nf3 Nc6 3.Bc4 Bc5 4.b4
FENr1bqk1nr/pppp1ppp/2n5/2b1p3/1PB1P3/5N2/P1PP1PPP/RNBQK2R b KQkq - 0 4
Of the candidates compared, Bxb4 scored highest (±0.00).
Searched with no candidate list imposed, at the same depth, the engine also chose Bxb4.
Tap a candidate move to change the board
| Move | Evaluation | Diff vs top candidate |
|---|---|---|
| Bxb4 | ±0.00 | baseline |
| Be7 | −0.21 | −0.21 |
| Bb6 | −0.31 | −0.31 |
| Nxb4 | −0.45 | −0.45 |
| d5 | −0.53 | −0.53 |
Position 4｜The poisoned pawn (Black to move)
1.e4 c5 2.Nf3 d6 3.d4 cxd4 4.Nxd4 Nf6 5.Nc3 a6 6.Bg5 e6 7.f4 Qb6 8.Qd2
FENrnb1kb1r/1p3ppp/pq1ppn2/6B1/3NPP2/2N5/PPPQ2PP/R3KB1R b KQkq - 2 8
Of the candidates compared, Qxb2 scored highest (−0.40).
Searched with no candidate list imposed, at the same depth, the engine also chose Qxb2.
Tap a candidate move to change the board
| Move | Evaluation | Diff vs top candidate |
|---|---|---|
| Qxb2 | −0.40 | baseline |
| Qc7 | −0.63 | −0.23 |
| h6 | −0.69 | −0.29 |
| Nbd7 | −0.72 | −0.32 |
| Be7 | −1.29 | −0.89 |
Where the difference comes from
In position 1 the engine’s line is the whole answer. After Nxd5 comes 6.Nxf7 Kxf7 7.Qf3+ — the Fried Liver, sacrificing a knight to drag the king out. With the d5 knight under fire, Black never gets time to make the king safe. The natural-looking h6 runs into the same 6.Nxf7 and scores lower still. The move that came out best, Na5, simply gives White’s strongest piece something to worry about.
The other three positions go the other way. In the Queen’s Gambit, dxc4 is 0.11 behind the top candidate — nothing. And in the Evans Gambit and the Sicilian poisoned pawn, taking is the top candidate outright.
Exceptions
Given that, "never take the offered pawn" is not a rule these positions support. Most gambit pawns can be taken and held. Of the four positions here, taking was a real mistake in exactly one.
What makes position 1 different is that White’s knight and bishop already point at f7 together. The question there was king safety, not material. The habit worth keeping is not "don’t take" but check, before taking, whether anything is aimed at f7 (or f2, playing White).
What this does and does not show
- It shows: what the engine returned for these 4 positions, with the chosen candidates searched to depth 20. The numbers are the raw output; none of them were adjusted afterwards.
- It does not show: anything about "all positions" — a sample of 4 is a very small slice of chess.
- The assumption: an evaluation assumes best play from both sides. Over the board, an inaccurate reply changes everything.
- Reproducing it: every number here is built from chess-analysis.json, the raw output of the run — positions, candidates, scores and principal variations. Running the same engine with the settings above reproduces them exactly: the search is single-threaded, the hash is cleared before each position, and the depth is fixed. A different engine version or hash size will give different numbers.
What to remember at the board
- "Never take the offered pawn" held in only 1 of the 4 positions
- In the Evans Gambit and the poisoned pawn, taking was best
- The Fried Liver is the exception, and it is about king safety
- Before taking, check what is aimed at f7 (or f2)
Try it yourself
You can try the same idea in a real game. Every move is scored by the engine, so the difference between principled and careless moves shows up immediately.