Entry 19

Dungeon and field generation, reproduced and checked against the game code

2026-09-22worlddecomptoolingtest

Filestools/dungeon.pytools/field.pytools/test_dungeon.pytools/test_field.pytools/eemu.pydocs/engine/dungeon.mddocs/engine/field.md

11 took three keywords to an area's fieldSeed, dungeonSeed[3] and attributes. This entry takes those seeds to the actual dungeon layouts and field terrain. tools/dungeon.py and tools/field.py reproduce the game's generators, and both are checked the way 11 was: by running the game's own code in tools/eemu.py and comparing every field it writes. The algorithms are on the dungeon page and the field page.

Dungeons

WORLD_MAN::GO (INF SLUS_202.67:0x0019f8e0) seeds the area RNG with dungeonSeed[game.dungeon] and resets randcnt, and DUNGEON::Generate (INF gcmn.prg:0x005c12a0) grows each floor from a medium, small or large room at the map centre: rooms spawn through their exits, each new room drawing its exit count and sides, until the floor has fieldrand(5) + roomMax rooms; a floor that stalls, or ends with fewer than two dead ends, starts over without rewinding the RNG. The down stairs go in a dead-end room that passes a 9% roll, the up stairs in room 0, and a dead-end medium room on the last floor becomes the Gott statue room. Every room then picks a model from a table by its exits and size, and every item-box dummy in that model rolls to be kept - which means each floor's layout depends on how many item-box dummies the earlier floors' models happen to contain. dungeon.py counts them from the models themselves, following the ExtObj indirection of 10.

WORLD_MAN::SetDungeonTypeFromField (0x0019cf50) picks the dungeon type, and so the CCS set (sd1 ... sda), from the field type, with overrides for some story areas. Story dungeons are not generated at all: MakeRealMap lays them out from ROOMDATA tables selected through EditDungeon.

Checked in eemu, with only asset loading and drawing stubbed: 300 random dungeons, 0 mismatches - 150 from random keyword triples on all five servers, 150 from raw inputs across volumes 1-4 including the forced-four-floor word 131 - covering 1,047 floors, 10,952 rooms, 759 floor restarts, all ten dungeon types and 297 statue rooms. Every floor's map, stairs, room models, rotations, positions, minimap codes, start orientations, statue fields, every kept gimmick, and the final RNG state and count are compared. All 88 EditDungeon layouts over 10 floors, 0 mismatches; all 480 inputs to SetDungeonTypeFromField, 0 mismatches. The item-box counts fed to the game code come from dungeon.py's own counting, so the counting itself is only checked through the layouts it produces.

Fields

WORLD::Init burns a fixed number of draws on snow and smoke (measured for all 110 field type and weather combinations), then WORLD::Generate (INF gcmn.prg:0x005a6da0) lays white noise from fieldrand(64), raises one large and three to eight small fractal hills (a 5x5 grid of random keys, spline-interpolated), places the dungeon entrance and a lake, scatters key, sub, base and tree objects with per-object placement tests, lays ground cover and picks the start. The hills are floating-point, so tools/eemu.py now interprets the EE FPU - including its non-IEEE rules: no denormals, no infinities or NaNs, saturation on overflow, truncation toward zero, and madd rounding the product first.

Checked in eemu against the real WORLD::Generate, comparing the height map bit for bit, the four chip grids, every object, every cover tile, the start position and the RNG state: the first story area identical, the unit tests over five field types, and a bulk run of 150 cases, 0 mismatches - 105 story areas and 45 random keyword triples, every field type but 4 (which has no field), 15,171 objects in all.

The first story area, generated

Bursting Passed Over Aqua Field (11, event 14): its dungeon is the hand-made D0001, 2 floors of 5 rooms; its field, from fieldSeed 1420855, is type 10 with 6 hills (heights 0-766), a lake, 107 objects, 1,550 cover tiles, and the dungeon entrance within 8,000 units of the start as that area requires. The outputs are in work/infection/areas/.