How to read a shogi engine evaluation (and why it is not a chess centipawn)

A chess player opens a shogi game record, runs the analysis, and sees the graph spike to +1800. Out of chess habit they divide by a hundred and read eighteen pawns. Two moves later it is back to +200, and they conclude the engine is broken or the position is insane.

The engine is fine, and the position probably is too. The division was almost right, which is the trap: a YaneuraOu pawn is 90 points, so +1800 is nearer twenty pawns than eighteen. What does not carry over is everything the number implies: how likely that edge is to hold, and how far a single capture can move it in a game where captured material comes straight back.

Nobody hands you a conversion table, so here is one. The reference comes first, and the argument about what an evaluation is for sits at the bottom.

What is a shogi engine evaluation, exactly?

It is a single signed integer that the engine emits at the end of a search, in units it defined for itself. Under USI, the Universal Shogi Interface adapted from chess’s UCI, the engine prints a line containing score cp <n> for a positional evaluation or score mate <n> when it has found a forced mate.

Two conventions matter before you read a single digit.

The sign is from the side to move. USI reports the score from the perspective of whoever is on move in the position being searched. Almost every graphical front-end flips it so that a positive number always means Black (先手, sente, the first player) is better. If your numbers seem to alternate sign every ply, you are reading raw engine output and not a normalised display.

The scale belongs to the engine. YaneuraOu, an open-source shogi engine, inherits the material table computer shogi has used since the Bonanza and Apery lineage: a pawn is worth about 90, a rook about 990. The NNUE network that produces the evaluation departs from pure material immediately, but the scale stays anchored there. That is close to a chess centipawn, 90 units per pawn against 100. Shogi evals look like four-digit numbers and chess evals like small decimals only because of display convention: chess front-ends divide by a hundred before drawing anything, and shogi front-ends print the integer. Tunings shift these constants, so if you need them exact, read them in the source instead of trusting a blog.

What do the evaluation bands mean for a club player?

Japanese broadcast coverage has settled on a rough set of bands, and they translate well to a club player’s decisions. The win-probability column comes from the logistic curve computer shogi uses to convert an evaluation to a win rate, 1 / (1 + e^(-eval/600)). The 600 comes from win-rate fitting in the elmo and YaneuraOu learning code; it is a tuning constant, not a law.

Two caveats matter more than the table does. The boundaries are conventions: a different engine, network or search budget draws them elsewhere, and nobody has a principled reason to prefer 400 over 450. And the win probability is a machine’s. Between two 1200-rated club players a +1500 position gets thrown away often enough that the number is advice and not a forecast.

Why do shogi evaluations swing harder than chess centipawns?

Because a captured shogi piece stays in the game. It goes to your hand, the mochigoma (持ち駒) that sit on the wooden stand beside the board called the komadai (駒台), and from there it can be dropped onto almost any empty square as an entire move.

A capture is a subtraction and an addition in the same move, and the evaluation moves by roughly the sum of both. Using the default material constants:

The bolded rows hold the most useful piece of shogi arithmetic: promoted pieces revert when captured. A tokin is worth a gold to the player who owns it and only a pawn to the player who takes it. That asymmetry is why building tokin is such a cheap, strong plan, and why losing one hurts you six times more than it helps your opponent.

Put the numbers together and winning a silver cleanly moves the evaluation about 990 points, which on the same scale is eleven pawns. In chess, winning a bishop moves you three pawns and you are pleased. In shogi, winning a silver drops you from “level” into the “winning” band in one move, and the engine is reporting the game correctly when it shows that.

The other half of the answer is what does not happen. Chess evaluations settle because material leaves and positions simplify toward technically drawn endings. Shogi has no such drain: trades load both komadai and sharpen the position. Shogi also has no stalemate draw and no draw by insufficient material. The only two draw mechanisms are sennichite, the fourfold repetition that gets replayed with colours swapped, and the rare double-entering-king impasse called jishogi (both are laid out on Wikipedia’s shogi page). An eval near zero in shogi means “unclear”, not “heading for a draw”.

Side by side:

Why is a piece in hand worth more than the same piece on the board?

The material table gives one value per piece, but the two states are not equal, and the NNUE network, which is not a material table, prices the difference where the table cannot.

A piece in hand has no travel time. A knight on your second rank needs several moves to matter; a knight in hand can arrive on any legal square in one move.

It has not committed to anything. A silver on the board is an argument about one part of the board. A silver in hand threatens every weak square at once, and your opponent has to answer all of them.

It is not blocking your own pieces, which counts for a lot on a 9x9 board where all nine of your pawns start on one unbroken rank.

The effect is strongest for the two pieces that can only move forward. A lance or a knight stuck behind your own lines is close to dead weight, since, as the piece guide explains, neither can ever turn around. The same lance in hand drops deep in enemy territory, aimed at a castle. The evaluation reflects that change even though the material table does not.

The drop restrictions are worth memorising, because the engine has already accounted for them and you have not: no two of your unpromoted pawns on one file (nifu), no pawn drop that delivers immediate checkmate (uchifuzume), and no drop onto a square where the piece would have no legal move.

What does a mate score mean in shogi?

It means the engine has stopped estimating. score mate 7 says “there is a forced mate, and here is the proof, seven moves long”, which is a different statement from “very good”. Three consequences follow.

It is off the eval scale. Mate in 11 is not worse than mate in 3; both are 100%. Engines store mate as a sentinel near the top of their integer range, so a bare five-digit figure in a front-end is a mate flag leaking through as a number.

A negative mate score is the one to read. score mate -5 means you are being mated in five, and it is the most instructive line any shogi engine hands you, because it usually starts with a drop you never considered.

An engine mate is not a tsume. Tsume-shogi is a strict problem form: every attacking move must be a check, the defender resists as long as possible, and every piece in hand must be used. Xiangqi has the same form without pieces in hand, the consecutive-check mate (连将杀); here is how Xiangqiful checked 902 of them. An engine’s forced mate has none of those constraints. Nor is there any score for hisshi (必至), the brinkmate where mate is unavoidable next move but not yet forced by checks, and club games are decided in that gap constantly. The engine shows you a slightly worse number and never mentions that the game was over.

How much depth does a club player actually need?

I have argued the general case elsewhere: engine depth at club level should be tuned down, not up, because analysis deeper than a player can calculate produces lines that are correct and unteachable. Shogi strengthens that argument in two ways.

The first is cost. With roughly 80 to 90 legal moves in a typical position against chess’s 35, each extra ply in shogi costs more and buys less. A shogi “depth 20” and a chess “depth 20” can’t be compared, because no honest conversion exists.

The second is what the extra plies contain. In chess, deeper search mostly surfaces quiet positional moves whose payoff arrives ten moves later. In shogi it mostly surfaces longer drop sequences, such as five-move nets that start with a pawn dropped on an empty file to be promoted three moves later. A club player had no chance of finding those at the board and no chance of generalising from them afterwards. The mistakes that actually decided your game are visible far shallower: the hanging silver, the knight drop that would have forked king and rook, the reflexive promotion, the edge pawn that opened your mino castle.

So this post gives no ply number, unlike the chess post does for Stockfish. The right depth moves with the position, the hardware and how full the komadai are. Search deep enough to prove the mistake, then stop.

Where can I get a free shogi engine evaluation?

Lishogi is free, open source and excellent. It runs online play, gives you a server-side analysis that searches harder than any phone will, produces a shareable URL for the game, and, because it is open source, lets you read exactly what the analysis did instead of trusting a description of it. If what you want is the number, on a board you can send to someone, lishogi is the answer and it costs nothing. PiyoShogi is also a good free app, with an analysis function and finely graded opponents. The honest guide to shogi apps for iPhone covers the rest of the category, 81Dojo included.

What an evaluation doesn’t tell you

Once you can read the number, it still doesn’t tell you what to do. -980 is a verdict: you are losing, and roughly by how much. It does not say why, which of your last thirty moves caused it, or what to practise tomorrow. An evaluation graph with a cliff in it is a crime-scene photo with no detective.

Turning the verdict into a sentence is the whole job of a training app, and it is harder in shogi than in chess because the chess vocabulary does not carry over. A chess trainer has fork, pin, skewer, back-rank, discovered attack. Shogi needs categories that chess’s taxonomy lacks: the drop that would have forked, the promotion that cost you a retreat, the castle brick you removed yourself, the mating race you lost by one tempo. Shogiful sorts mistakes into 12 types written for shogi, among them Drop Mistake, Promotion Mistake, Castle Weakness, Missed Tsume, Hanging Piece, and Endgame Error. Each mistake gets a card with the move you played, the best move, the cost in pawns, and a coach’s tip, written from a template for its type, and each day’s training is built around the weakness that keeps recurring. YaneuraOu’s NNUE network runs on the device, so the review needs no server and no account. Free covers two game analyses, three training puzzles, and two coach hints a day. Premium is $1.99/month, $9.99/year, or $19.99 once; it removes the daily limits and adds Deep Lines, the engine’s top three lines for any position.

That design has a cost. Shogiful leads with the sentence and puts the number second, and it has no online human play at all, so if you wanted the eval graph itself, deep and shareable, lishogi does it better and for free. Chessful makes the same trade on the chess side with Stockfish, which is how I know it is a real trade and not a rationalisation: both apps give up the raw number as the headline in exchange for a sentence you can act on.

TL;DR

• A shogi eval is not a centipawn, but the units are close enough to fool you. YaneuraOu anchors to a table where a pawn is about 90 against a centipawn’s 100, and shogi front-ends print the integer without dividing. Divide by roughly 90 to get pawns, then stop trusting the comparison: ten pawns is a decided chess game and an 82% shogi position.
• The bands for a club player: under ±150 level, ±150-400 a small edge, ±400-900 a clear advantage, ±900-2000 winning if converted, beyond ±2000 decided. These are display conventions, and they move with the engine.
• Every capture counts twice, because the piece goes to your hand. Winning a silver swings the number by about 990, eleven pawns’ worth. Promoted pieces revert when captured, so a tokin is a gold to its owner and a pawn to its taker.
• In most positions a piece in hand beats the same piece on the board: it has no travel time, no positional commitment, and it threatens every weak square at once. This is most extreme for the lance and knight, which cannot turn around.
• A mate score is categorical. Mate in 11 is worth exactly as much as mate in 3, and a negative mate score is the most instructive output a shogi engine produces. An engine mate is not the same as a tsume, and no engine reports hisshi at all.
• More depth doesn’t make analysis more instructive. Shogi’s branching factor makes plies expensive, and what they buy are long drop sequences a club player could never have found. Search deep enough to prove the mistake, then stop.