Rules

gen-generals-io — the rules

Four commanders. One 16 × 10 grid. Everyone starts on a crown in a corner, sees only what they own and the ring around it, and spends one move a turn pushing armies outward.

The board

• boardW × boardH = 16 × 10 (12 × 8 in blitz), generated once at reset from the episode seed and four-fold symmetric: the top-left quadrant is drawn, then mirrored into the other three. Every seat gets a congruent start — the same crown offset, the same mountains, the same cities and the same distances to each rival.
• Cells are plain, mountain, city or general. Mountains are impassable and never owned. A cell holds an integer army.
• 8 mountains and 2 cities per quadrant (32 and 8 on the board), placed by rejection sampling at Chebyshev ≥ 2 from the crown (mountains) and ≥ 3 from the crown and ≥ 2 from each other (cities). A symmetric connectivity repair then opens any pocket the mountains sealed off, converting the first mountain on the cheapest path and its three mirror images to plain — so the layout stays symmetric and every non-mountain cell stays reachable.
• Neutral cities hold 40 armies (50 in citadels) and are the only toll on the board.

The turn

One tick is one turn. maxTurns = 240 (160 in blitz).

1. Directive install — every 8 turns each living seat's plan is installed.
2. Move compilation — the deterministic captain compiles one plan into at most one move per seat, from that seat's fogged view only.
3. Move resolution, one at a time, in the rotated order order[k] = (turn + k) mod 4 — priority rounds the table every turn.
   • A move is discarded (and counted in invalidMoves) if the source is not owned, holds fewer than 2 armies, the target is off the board or a mountain, or the seat is no longer alive.
   • amount is clamped into 1 .. army(source) - 1: a tile always keeps at least one army.
   • Onto your own tile the armies add. Onto a neutral or enemy tile: send more than the garrison and you take the tile with the difference on it; send the same or less and the tile keeps the difference and its owner.
4. Per-turn growth — every owned city and crown gains +1. Neutral cities do not grow.
5. Periodic growth — every growthPeriod (25) turns, every owned cell gains +1, whatever its kind.
6. Vision and memory — a seat sees a cell if it owns it or the cell touches one of its cells (all eight directions). What it saw is remembered with the turn it was seen; the army on a remembered cell is unknown.
7. The sim guard runs every turn.
8. The hash is written and the end conditions are checked.

Crown capture

Move more armies onto a rival's crown than sit on it and you take everything they own: every other tile they hold becomes yours with its army halved (integer division), their crown becomes a city of yours holding the surviving attacker army, and they are out of the game. The captor gets a generalcaptured event and the victim an eliminated one.

Scoring

Measured at the final turn, first difference decides:

1. still alive
2. among the dead, who lasted longer
3. most land
4. biggest army
5. most cities

placePoint(rank) = (4 - 1 - rank) / (4 - 1) → [1, 2/3, 1/3, 0], averaged over a tie group. sum(scores) == 2.0 on every episode, ties included: strictly constant-sum, higher is better, nothing is ever negative.

Endings

No other string is ever emitted.

The scripted baselines

Both emit the same plan object an LLM does, through the same validator, and both are pure functions of the seat's own fogged view.

sprawl — defend when the crown is threatened, expand while land is under a quarter of the board, attack at the nearest visible enemy tile otherwise. reserve 0, cities cheap, scouts 1. It is also the per-turn fallback, the driver of a no-show seat, and the default.

crown — defend when threatened, raid the nearest known crown, expand otherwise. reserve 20, cities never, scouts 2. Deliberately a different SHAPE, so the ladder gets a spread rather than two versions of one bot.

Documented divergences from generals.io

This is an adaptation of a public specification, not a reproduction of anyone's engine, and no test here compares a trajectory to a reference implementation.

• A move carries an explicit amount in 1 .. army-1 rather than only "all" or "half". A strict superset, and it is what makes reserve exact.
• Priority rotates by turn number rather than by a server-side move queue.
• Generals start at 1 army, on a board generated by this repo's own symmetric generator.
• Swamps, lookout towers, desert tiles and the 50/50 city-spawn variance are absent.
• A 240-turn clock stands in for generals.io's open-ended games.