The Riverbank Cannon Problem

The rush is invisible

The b-file route announces itself: Black’s cannon watches that file and sees something land on the riverbank. Red does not have to go that way. Up the empty d-file, nothing Black owns sees a single point: d3, d5, e5, mate on move 4. The only warning is that a red piece left home, which describes every game ever played.

So the defense cannot wait for a warning; it has to be played every game. The rest of the article draws the visible route so the diagrams read easily. The threats are the same either way.

The natural defense is a landmine

There is exactly one screen between cannon and general: Black’s own center soldier. A cannon needs exactly one, so any second body on the center file kills the mate. The natural pick is advisor up. It seals the center and nothing else, and it plants a mine: Red slides to an elephant wing and takes.

There is a save: play the poisoned wing’s elephant to the middle immediately, and both grabs die. But that is the move you should have opened with. The advisor spends a whole move doing a fraction of the elephant’s job, and while you fix it, Red takes a rim chariot for free.

One elephant move holds everything

The move that works is the standard developing move of xiangqi: elephant to the middle.

Snipe dead, both elephant wings covered, one move. Xiangqi theory reached this square centuries before the fog did. A cannon to the same point works too.

The chariot gamble

That leaves the edges: from the riverbank corner the cannon shoots the chariot through the edge soldier, and the elephant does nothing about it.

Move order decides this. The mate cannot arrive before Red’s third move, so the elephant is still in time on Black’s second. The edges cannot wait: Red picks his corner on his second move, and a soldier pushed after that is too late. So soldier first, elephant second. The pushed soldier watches the one point the cannon must fire from, and eats it on arrival.

That leaves one honest gamble. Red commits blind to a corner on move two: half the time it is the watched one and he trades his cannon for a soldier, half the time he wins a chariot. Played out by the engine, the branches roughly cancel. One fair coin flip per game, and only if Red commits immediately: given a third move, Black closes both edges.

When the cannon dies

Half the flips, Red loses the cannon to the soldier. He recaptures the soldier (the engine did, in every playout), and the recapture is a gift: it is now the screen, and Black’s cannon shoots the corner chariot straight through it. The engine scores this -0.55 for Red: attack over, a cannon down, nothing to show.

Can Red defend the shot? Two ways, neither happy. Declining the recapture keeps two screens on the file, but the crossed soldier just keeps eating and the shot reopens a move later, one soldier cheaper. Recapturing and then blocking with the horse holds the chariot. In six engine playouts from the fired tripwire, the whole sequence played itself out, capture, recapture, counter-shot and all, and Black converted five (the Tripwire games in the companion study). The block:

When the chariot falls

The other half, Black is a chariot down. The engine scores positions from -1, lost, to +1, won; it calls this one -0.75, which is roughly one win in eight. Close to lost, not over. And the landed cannon looks scarier than it is: every capture it has loses on the spot to a recapture. Its real power is the freeze: the horse and the elephant beside it are holding the back rank shut, and if either ever moves, the next shot is the general.

So how does Black fight? The engine showed three ideas. First, patience: touch nothing on the back rank. Second, an immediate simplification: in our first playouts its opening move from here, every game, was to snipe Red’s home horse with its own cannon, over Red’s home cannon as the screen, trading cannon for horse to blunt the attack. Third, counterattack on the other wing, where your own cannons still have a game; the next section shows the sharpest version. And it escapes: of the ten forced-grab games, Black won two outright (Rim gambit games 2 and 8 in the companion study, in 84 and 56 plies) and held three more to the ply cap. One win in eight is the eval; in the decided games, the practice was one in four.

Whoever moves the wall first dies

One pattern decided games in every branch, and it travels to any fog game with long-range pieces: a cannon on its firing point is also the block against the enemy cannon opposite. Whoever steps away first, the other fires through the hole. And the fog baits you to step away: from the wall, an enemy soldier looks free. Take it and you are mated on the reply.

Which answers a fair question: why not go to the riverbank yourself as Black? You can, and it punishes any Red who grabs. Red has exactly one sound reply: seal his own center before touching anything. Then the battery never fires.

What the engine says

I forced the rush onto the board and let Misty play both sides, thirty games: twenty through the snipe attempt, ten through the chariot grab. The snipe never fired once; Black sealed within three moves in eighteen of twenty. The rush games went Black 9, Red 4, seven ply-cap draws; the forced grabs went Red 5, Black 2, three caps. Then I made Black answer the same forced rush with this article’s exact line, eight more games: Black won six, no draws. The line does not just suit humans; it outscored the engine’s own defensive choices. Here is one of those games in full:

All the games behind this article are in the companion study: the theory lines, twenty forced-rush games and ten forced rim-gambit games with the engine on both sides, and sixteen free self-play games. Flip the study board to Black’s fogged view and step through what he actually saw.

Left to choose freely, sixteen more games, Red won eleven of the twelve decisive and never once played the rush. Lean samples, so treat the counts as direction. The position evaluations are firmer ground: they hold still under a sixteen-fold compute increase.

One last check: I forced Black through the line this article teaches against a free Red, eight games. Black won five of the seven decisive, against one of twelve when choosing freely. The three insurance moves cost nothing.

Average the flip branches and the committed rush is worth about +0.10 to Red, the same ballpark as the +0.06 he starts with: the whole first-move advantage, spent in one gamble. Initiative in fog is real, probably bigger than in chess, because defense is paid blind. The rush is the loudest way to spend it, and the loudest way is answerable.

The line to learn

As Black: edge soldier, elephant to the middle, horse to the other edge. After three moves the only target left is the chariot on the edge you did not push, and only if Red committed on move two. Never seal with the advisor. And before you move any piece near a landed cannon, count screens: most losses we found came from moving a piece that was quietly holding a firing line shut.

As Red: the rush beats anyone who has not read this far, and it is even money against anyone who has. If you play it, commit to an edge on move two or not at all, and seal your own center before cashing any grab: every scripted Red that grabbed first got mated.

Step through it yourself

Every line and every game in this article is in the companion study, on a board you can flip to either side’s fogged view. When you are ready, Misty will punish you while you learn the line. No account required.