Rules

lux-ai — the rules

A port of the Lux AI Challenge Season 1 rules onto a 16 x 16 mirrored grid. Two sides gather wood, coal and uranium, spend research to unlock the better fuels, grow cities out of city tiles, and every ten turns the sun goes down and each city must pay its light bill or die on the spot. Most city tiles standing at turn 360 wins.

The clock

• One tick is one Lux turn. maxTurns is 360 (duel, scarcity) or 200 (skirmish).
• A cycle is cycleLength 40 turns: turn mod 40 < 30 is day, turn mod 40 >= 30 is night. Nine full cycles in 360 turns.
• A directive turn is every directiveEvery = 10 turns, beginning at turn 0: 36 directive turns per episode. Ten divides the 30/10 cycle exactly, so a seat is always asked for a new directive on the turn the sun sets.

The board

mapSize is even and is 16 in the default variant. The map is generated from the episode seed and is perfectly mirror-symmetric about the vertical midline; seat 0 owns the left half, seat 1 the right. Symmetry is the integrity pin: neither side can be dealt a better island.

The turn, in order

Every turn runs these steps and nothing else mutates the world. An action that is illegal at the moment it is evaluated is discarded and costs no cooldown.

1. Directive install (directive turns only).
2. Order compilation — the deterministic micro layer emits at most one action per unit and per city tile whose cooldownTenths == 0.
3. City-tile actions, ascending tile index, seat 0 then seat 1: research (+1 point), build_worker / build_cart (only while units < cityTiles). Any accepted action costs the tile 100 cooldown tenths.
4. Transfers, ascending giver unit id, to an orthogonally adjacent unit of the same team.
5. City building, ascending unit id. A worker on empty terrain with no city tile and wood + coal + uranium >= 100 spends exactly 100 resource units cheapest-first and plants a tile. Two opposing workers on one cell: neither builds. Two friendly: the lower unit id builds. The new tile merges into the lowest adjacent same-team city id, fuels summing; road level 6.
6. Movement — illegal targets discarded; a monotone blocking-and-contention fixed point decides who actually moves; survivors apply simultaneously; a cart paves the empty cell it lands on.
7. Resource collection — only workers, wood then coal then uranium, over every tile of that kind in ascending index. Rates 20 / 5 / 2 per adjacent worker per tile; when a tile cannot pay everyone, the amount is split evenly with the first amount mod |M| workers taking one extra. Coal needs 50 research points, uranium 200.
8. Deposit — every unit on a friendly city tile empties its whole cargo into that city at 1 / 10 / 40 fuel per unit, every turn, day and night.
9. Night burn (night turns only). Cities first: upkeep = 23 * tiles - 5 * adjacentPairs; a city that cannot pay is destroyed entirely. Units second: a unit on a surviving friendly city tile pays nothing, every other unit burns 4 (worker) or 10 (cart) fuel from its own cargo, and one that cannot pay dies.
10. Cooldowns — cooldownTenths -= 10 + 2 * road; city tiles -10.
11. Wood regrowth — every wood tile with 0 < amount < 500 grows by max(1, amount div 50). A tile mined to exactly 0 never comes back.
12. Sim guard — checkLuxInvariants().
13. Hash and end check.

Winning

Measured at the final turn, first difference decides: more city tiles, then more units, then more banked fuel, else a tie. The league score is the match point: 1.0 / 0.5 / 0.0, and scores[0] + scores[1] == 1.0 on every episode.

Documented divergences from Lux-Design-S1

This is an adaptation of a public specification, not a reproduction of the JavaScript engine. No test compares a trajectory to a reference implementation.

Two amendments to the design note's own rule order

Both are recorded here because they change how a game plays, and both exist because the literal reading produces a dead game:

1. Production before research. The design note orders the city-tile rules research-first. S1's unit cap (units < cityTiles) makes that self-defeating at the opening: a side with one tile and one worker spends every tile action of the first hundred turns on research points and starves, because a research point does not mine wood and the night bill does not wait. Research therefore YIELDS while the side is both under the unit cap and below its directive's unit target — the exact trade the system prompt describes to the model.
2. prospector carries forester's night guard, and plants a seed blob first. The note's prospector researches to 200 before it expands; with one city tile earning one point every ten turns that is 2000 turns, so it never moves. It now expands to prospectorSeedTiles tiles first and, like forester, drops to stance: "fuel" whenever one of its cities is inside foresterFuelNights nights of starving. Without the guard it is wiped out in night 1 on most seeds, which is a dead filler rather than a control.

The scripted baselines

forester (PLAYER_SCRIPTED=forester) — the certification player, the per-turn LLM fallback, the driver of a no-show seat, and the published default. Wood until coal is researched, expand unless a city is inside foresterFuelNights nights of starving.

prospector (PLAYER_SCRIPTED=prospector) — the second filler, deliberately different in shape: it buys the fuel ladder early and pays for it in tiles. forester beats it at the pinned seed, which is a real bar for a champion to clear.

Both emit the same directive object an LLM does, through the same validator, and neither ever writes a note. Their tuned numbers are the pick from tools/tune_baselines.nim, pinned in tools/ci/baseline_tuning.json.