SE!ZE was made by James Wallis for Chivas Regal and Charles Leclerc. There is no published rulebook. This is a program that plays it anyway, and a record of which parts we had to invent.
Those fourteen include whether captures work diagonally, what happens when
a piece lands on a second power token, and whether the game has any draw
rule at all. We picked values so the engine would run. They are marked
ours in the ruleset and they are probably not all right.
Seize is French for sixteen. Leclerc's car number is 16 and he is Monégasque. The game has sixteen pieces, and they start on the sixteen squares of the middle 4×4.
Alpha-beta search with a transposition table, late move reductions and null move pruning. It reaches depth 9 on a laptop in about a second. The evaluation has twelve terms, fitted by logistic regression against the results of 2,500 self-play games. 74% of those games finish with a winner.
That last number was 10% two days ago. Almost all of the difference came from fixing three bugs rather than from any new idea. A repetition check was comparing against the wrong player's positions, and two invented stopping rules were ending games somebody was still winning.
The search and the evaluation are ordinary work and they are measured: node counts, branching factors and win rates on this page come from runs that can be repeated, and the ones that failed are written down next to the ones that worked.
The rules underneath them are a different matter. Nobody outside the publisher knows them, this project included. The numbers mean "given these rules, this is what the engine finds". That is a real result. It is not a claim about how SE!ZE is actually played.
A published game. If a real match contains a move this engine calls illegal, that points at which of the fourteen guesses is wrong. One of those is worth more than any amount of further search.