A turn is a move plus a duck placement
Every turn you move a piece and then place the duck. The move is xiangqi and you already know how to think about it. The placement is the new game, and it is where first games are lost.
The duck is a cannon screen for whoever moves next
A cannon captures by jumping exactly one piece, and the duck counts as that piece. This is the most important thing on this page, and it follows from the order of a turn rather than from any subtlety: you place the duck at the END of your move, so the next person to play is your opponent. Any cannon line you just completed is theirs to fire down first.
The practical habit: before you place the duck, look along the ranks and files it will sit on and check whether an enemy cannon has just been handed a target. The square that most obstructs your opponent is usually a square next to their pieces, which is exactly where it is most likely to arm one of their cannons.
There is no check and no checkmate
You may leave your general attacked, move it onto an attacked point, and nothing on the board or in the interface says a word. Your opponent wins by actually taking it, and the game ends there.
This is not a theoretical risk. All seven of the engine games we published end with a general captured outright. Not one reached a resignation, a repetition or a stalemate: every game ended because somebody could take the general and did. Count the threats yourself, every turn, including the ones the duck just opened.
The duck must move every turn
A duck jamming a horse’s leg, filling an elephant’s eye or holding a file is doing that job for exactly one turn. Then it is your problem again: on your next turn you have to pick it up and put it somewhere else.
So a block is never a structure, only a delay, and it is worth one turn of inconvenience rather than a wall. It also means you can be forced to open a line against yourself, because standing still is not among your options.
The bot places the duck at random
The bot is a patched Fairy-Stockfish and it plays the piece half of its turn properly. The duck half it essentially guesses, and that is measurable rather than a suspicion.
We scored every legal duck placement after the engine’s own opening move, 58 of them, and the engine returned an identical evaluation for all 58. Raising the search from 300,000 nodes to four million changed the number but not the verdict: still identical across every placement. Fairy-Stockfish has no evaluation term for a duck or a wall, so nothing in its scoring can prefer one point to another, and the placement it reports is whatever its move ordering happened to surface.
That makes it a real opponent for the xiangqi half of the game and a poor teacher for the duck half. If its placements look arbitrary to you, it is because they are. The games are still worth reading for the piece play, and the study below has all seven.
The generals may face each other
In ordinary xiangqi that is forbidden. Here the prohibition stopped making sense once check was gone, so flying the general is simply a capture that ends the game. It is one more threat with no warning attached, and one more line the duck can open or close.
Games run 120 to 229 plies
That is the range across the seven engine games. Expect a full-length game rather than a novelty, and pick a clock accordingly: the two paces offered here are 5+5 and 10+5 for exactly this reason.
Play it
Eight engine strengths, or a friend. The rules page has the whole thing with diagrams, and the study has all seven engine games if you want to watch it played first.