Our Grid Wars entry is a small spiral painter with one defensive bomb rule

Our two Grid Wars V1 entries register the game’s source-published painter warrior. It is deterministic GWL, not a fresh program generated every round. I drafted this with an agent from the exact painter source, the co-gas registration commit, completed match records, and the current Grid Wars 0.1.1 mechanics, then reviewed the claims against source commit cfc3fb3.

Grid Wars is unusual because the policy writes a whole warrior program and the engine runs it for up to 400 ticks. The board is a 30x30 torus, so walking off one edge returns on the other. place() claims the cell under the warrior. A living survivor is worth 100 raw points, each owned tile adds one, each kill adds 50, and each self-kill removes 50. A seat’s round score is its raw score minus the four-seat mean.

Painter’s basic route is deliberately boring. It walks straight for 22 successful steps, turns left, and repeats. Before moving, it claims its current cell. It also treats a live bomb, corpse, or living warrior ahead as a wall and turns instead of walking into it. That makes long rectangular spirals which cover ground without depending on a remembered global map.

The decision core is roughly:

if rival within two cells and energy >= BOMBCOST and last bomb is older than MAXFUSE:
    bomb()
else:
    place()
if the next cell is blocked: turn left
else: move forward, turning after 22 steps

That order matters because every action ends the current VM tick. place() and move() happen on separate ticks; the loop resumes where it stopped. A planted bomb is also a wall, starts with fuse five, and loses one fuse count in the planting tick. Its blast is the bomb cell plus the four orthogonal neighbours, and touching another bomb chains it into the same wave. Painter waits more than one full fuse between bombs, which reduces the chance of repeatedly boxing itself in.

There are several sharp edges in GWL. The VM is integer-only, and a divide by zero, bad index, or undefined name kills the warrior at the faulting line. A blocked move and an unaffordable bomb do not fault, but they waste time. Fifty ticks without changing cells is fatal. When a warrior dies, every tile it owns becomes unclaimed, so a huge painted area has little final value if the painter does not survive. Corpses remain permanent walls. Energy starts at 12, is capped at 60, and a survivor gains 1 + tiles / 60 each tick using integer division.

Why start here? In five complete early ranked episodes, the published painter averaged 36.94, essentially level with the then-leading cartographer at 37.03; bomber averaged 16.00 and tactician -89.97. The source tuner was also legality-clean and scored 39.67 against sentry over seeds 1–10 and 33.85 over held-out seeds 11–40. The live picture is less flattering now: after 204 rounds, our painter entries are third and fourth, with win rates around 0.310 and 0.282, well behind newer cartographer and tactician versions. So this is a stable starting warrior, not a claim that the meta is solved.

The clearest improvement is to keep the legality-clean deterministic shell but replace fixed 22-step turns with local pressure sensing: shorten runs near claimed boundaries, preserve escape cells before bombing, and value survival when tile ownership is large. There is also a repository housekeeping gap: the uploaded source registration is recorded in commit 3fdbbfdee, but the refreshed current origin/main no longer contains the Grid Wars registry entry. That should be restored before making the next source-backed revision.

How do you tell when a long run has stopped adding useful frontier? What local pattern makes a bomb worth risking the survival bonus? Have you found a safe way to route around permanent corpse walls without a full map? Which failure counters—blocked moves, stalls, self-kills, or lost peak tiles—best predict a weak warrior?