Koshin StudioSetting the rules of Negaeri Shogi by playing AI against AI: one change that worked, six that broke
Development log of Negaeri Shogi / September 2026
Negaeri Shogi is a shogi-style game on a 6×6 board played a bit like Othello. Pieces you sandwich defect to your side and are promoted at the same time. If the new moves of a promoted piece create another sandwich, the flips continue. That is a chain.
The rules come down to about ten numbers. I did not set them by feel. I ran dozens of AI-vs-AI games per candidate and measured. Only one change really worked; most of the others broke the game. I keep the ways they broke here.
The numbers come from the measurement log in the game’s README (as of 13 September 2026). They are AI-vs-AI results. Balance between human players has not been verified enough yet.
What I measured, and how
The AI is a home-made alpha-beta search. I made depth-3 AIs play each other and counted:
- Share of drops that cause a flip. If low, most moves do nothing.
- Average game length. Too short means a one-sided game.
- First player’s win rate. Closer to 50% is fairer.
- Share of games with a chain of 3 or more. Chains are the point of the game.
- Lead changes. Too few: whoever is ahead early wins. Too many: it flips every move and feels random.
Candidates were compared at 40 games each; final numbers use 100 games. Anyone can re-run it with the README commands (npm run report, npm run sweep).
The first rules barely did anything
At first a sandwich only worked along the directions a piece can move in shogi. In AI-vs-AI games, only 2% of drops caused a flip. Games lasted 16 moves and the first player won 71%. A flipping game where almost nothing flipped.
| Stage | Drops that flip | Avg. moves | First player wins |
|---|---|---|---|
| Start (move directions only, empty board) | 2% | 16 | 71% |
| All 8 directions | 8% | — | — |
| Formation start position | — | 29 | — |
| Pieces next to the king are safe | 42% | 51 | 33% |
| One extra pawn in hand for the first player | 42% | 61 | 47% |
| Board edge counts as the end of a sandwich | 81% | 53 | 59% |
The first player’s win rate jumps 71%, 33%, 47%, 59% because each rule change moved the advantage somewhere else. Every change needed a fresh measurement.
The change that mattered: the edge is an anchor
If a run of enemy pieces ends at the edge of the board, it counts as sandwiched. Pieces pushed against the edge defect together. This took flipping drops from 42% to 81%. Compared with other ideas at 40 games each:
| Idea | Flips per move | Games with a 3+ chain | Lead changes |
|---|---|---|---|
| Before | 0.76 | 15% | 7.0 |
| Edge as anchor | 2.51 | 85% | 7.4 |
| + dropped piece is promoted too | 3.06 | 80% | 9.9 |
| Dropped piece promoted, alone | 0.72 | 13% | 5.0 |
| Chain limit of 3 | 0.71 | 8% | 5.8 |
| Piece-count result at 60 moves | 0.69 | 3% | 4.8 |
Promoting the dropped piece too adds flips, but lead changes rise to 9.9 and the lead flips nearly every move, which thins out the shogi side of the game. I left it out (one setting turns it on).
What I tried and dropped
| Idea | What happened |
|---|---|
| Off-direction sandwiches reach only 1 square | Games with a 3+ chain fell from 95% to 13%. The reach of a sandwich was the engine of chains. |
| 5×5 board | Flips per move rose from 2.88 to 5.99, but the first player won 70%. Games were short (35 moves). |
| Remove the king guard | Games ended in 17 moves on average and the first player won 95%. Guards defect at once and the king is left bare. |
| Chain limit of 8 | Almost identical (2.86 vs 2.88 flips per move). The limit of 5 hardly mattered. |
| Dropped piece promoted by a big flip | More flips, but 9.9 lead changes: it flips every move. |
| Piece-count result at 60 moves | Games with a 3+ chain fell to 3%. |
The most useful row was “a limit of 8 changes nothing.” Raising the limit sounds like it should add chains. Measured, the result barely moved, so the limit of 5 was not doing much. I was about to tune a number that did not matter.
Pieces next to your king never defect
In this game a piece next to your own king does not defect. Removing that rule is the third row above: 17 moves on average and a 95% first-player win. Guards defected in a single move and the king was taken early. It is one plain sentence in the rules, and the game does not work without it.
Current numbers (100 AI-vs-AI games, depth 3)
| Measure | Value |
|---|---|
| Average length | 52.7 moves |
| Board fill | 61% |
| Drops that flip | 81% |
| Flips per move | 3.08 |
| Games with a 3+ chain | 87% |
| Games with a 4+ chain | 38% |
| Average lead changes | 10.4 |
| Games with no lead change | 0% |
| Games where the loser led at some point | 98% |
| Point at which the result is decided | 81% of the way through |
| First player’s win rate | 61% |
The board changes hands again and again and stays close to the end, yet it is not luck: a depth-3 AI beat a depth-2 AI in 20 of 24 games (as first or second player). 98% of games end by taking the king, so sandwiching stays a means and the shogi goal stays the goal.
Still unsolved: the first player’s edge
In the final measurement the first player won 61% (59% in an earlier, separate measurement). That is 11 points from 50%. I tried five handicap setups and found none clearly better.
| Handicap | First player wins |
|---|---|
| None | 33% |
| One extra pawn in the first player’s hand (current) | 61% |
| One extra pawn on the first player’s side of the board | 38% |
| Move one of the first player’s board pawns into hand | 64% |
| Extra pawn in hand for the first player and extra pawn on the board for the second | 27% |
Pieces come in whole units, and the same “one pawn” is worth very different amounts depending on where it is placed: a point-symmetric addition moved the result by +5 points in one case and −34 in the other. Adding up small pieces cannot tune it to 50%.
In online play, matchmaking gives the first move to the lower-rated player, which partly cancels the gap. That is not a fix. Games between humans are still to be checked.
What I learned
Half the changes that felt like good ideas broke the game when I measured them.
- Build the measuring tool first. Being able to run AI-vs-AI games automatically is what let me compare six ideas under the same conditions.
- Some numbers do nothing. The chain limit gave the same result at 5 and at 8.
- Ideas that break the point of the game only show up in measurement. “Reach one square” looked like it would restore each piece’s character; it cut chains from 95% to 13%.
- Say what is not fixed. The first-player gap is still 61%.
This is the English version of the original Japanese article. Numbers come from the README measurement log (13 September 2026). Negaeri Shogi is free to play in the browser.
More in English