Open the books

RAW FILE: REPORT.md, as written during the jam

Source research/altbuild/REPORT.md in the jam repo. Last git commit of this file: 2026-11-15. About 531 words. Scrubbed only for private details and one line about another entrant. Mistakes left in on purpose.

SPOILED: built to the lens, three-room test (Sun Nov 15 2026, 08:56 to 09:10 ET)

Don asked for one Opus job to try building rooms from generated imagery, to compare with the rooms as built. This is the helper's report, condensed by the director.

Method

Raw frames at 2:52 (tank room), 5:44 (canteen) and 6:06 (lift) from the pkg3n take. Plates from FLUX.2 klein 4B edit on tower 1 with the game frame as the only reference: 2 MP, 4 seeds per room, 14 s each. Depth from MoGe-2. The move is a CPU per-pixel splat with a z-buffer over a filled background, 24 fps, 1080p. An enemy cut from the game frame was placed in each plate as a check.

Verdict per room

What broke

A whole 9 minute rail game this way

Could one room be replaced today

No. The only candidate was the tank room's back wall, and a card there would clash with the game's warm light and leave a seam at the floor.

What the director took from it

The plates beat the built rooms on WEAR, not on geometry: stains, wet floors, rust streaks, grime at the foot of walls, water-marked ceiling tiles. That can be put into the real rooms as world-aligned grime layers and streak decals, and was ordered at once (eng11). For a sequel: retest with clean plates, a proxy room and a matched light rig on a slow shot.

Files

sheets/ (comparisons, candidates, depth, keyframes), clips/ (dollies, plain and with an enemy), plates/, depth/, frames/, scripts/.