We had an AI play our board game 1,800 times

Here is the problem with balancing a dice game: the thing you care about most almost never happens. A run of SHIPLESS is a string of small gambles — a locked door, a thing in the dark, a crewmate who is almost still a person — and the number that matters, did you escape, turns on the tail of a distribution you'd need hundreds of plays to even see. No designer plays their game five hundred times. We didn't want to ship on a hunch.

So we built SHIPLESS around a pure, seeded rules engine from day one. Every roll, every draw, every point of damage flows through one deterministic function: reduce(state, action), with all randomness carried in the state itself. No hidden dice, no wall-clock, no Math.random. Give it the same seed and you get the same run, byte for byte. That single discipline is what makes the game legible to a machine — and once a game is legible to a machine, you can have the machine play it.

Skilled agents, not random flailing

A random policy tells you almost nothing; it dies in the doorway and never uses its gear. So we wrote skilled policies — agents that spend the CHARGE die when it's worth it, save their heals, and pick the fights they can win — and pointed them at the engine across all six classes and hundreds of seeds each. Then we read the aggregate: win rates, damage curves, how often anyone ever unjams a gun, whether the boss is a filter or a formality.

The first sweep corrected a number we'd been quoting for a month. The real skilled win rate wasn't 12.5%. It was 18.9%.

That's the small stuff. The agents also found two things that would have embarrassed us on a table.

The co-op collapse

SHIPLESS is meant to read 1–4 players, co-operative on the box. We modelled all eighteen enemy rows against real class dice at one, two, three and four seats — and watched the difficulty fall off a cliff. Average attrition dropped from 23.3% at solo to 2.1% at four players: an 11× collapse. Not one of the eighteen rows held its shape. Every extra survivor brought a whole new fistful of dice to cancel symbols, while the enemies barely scaled — so a four-player table was a cakewalk wearing a horror skin. You cannot print "1–4 players" over that.

Because the whole thing is content-driven, the fix was small and measurable: give each row a per-player scaling term, re-run the sweep, and land four-player difficulty back within about ten percent of solo. The model told us exactly how much to add before we cut a single card.

The most dangerous dial in the game

In-run progression lets you pull dice from a bag as you go, and each level appends a few enemy symbols to keep pace. How many symbols? We tried two rounding rules that look almost identical on paper. The difference between them was worth forty-five points of win rate. Forty-five. A rule you'd skim past in a rulebook was, quietly, the single most sensitive lever in the system — and we only know that because something played it a thousand times each way and reported back.

None of this replaces humans at a table — that's the next milestone, and it's how the rulebook gets proven. But it changes what the humans are for. They're not there to grind out rare events; a machine can do that overnight. They're there to tell us how it feels. The agents keep the math honest so the playtesters can spend their attention on the only thing that was ever really theirs: the dread.

SHIPLESS is free to play in your browser right now. The printed box comes to Kickstarter next. If you want the launch signal — the Steam release and the campaign — leave your email, or just go get eaten by the Erebus.