Show HN: I Turned DOOM into a Transformer

Doomformer is an experimental compiler that turns integer neural networks into playable ZDoom maps, then runs them in ViZDoom. Its central idea is to make game geometry and mechanics perform computation: duplicated player characters called voodoo dolls act as arithmetic units, floor lines encode weights, and sector lighting and floor or ceiling heights store intermediate values. Python reads those values from the running game and checks each against a separate NumPy integer implementation; any discrepancy, even one unit, fails the run. The system supports small multilayer networks and several transformer components, but operates within value-range constraints and is extraordinarily slow: a handwritten digit takes about 163 seconds of game time at the default schedule, or eight seconds with an optimized schedule.

The demonstrations include a marine drawing digits on a floor grid, a map that computes a digit classification, and a museum level where changing one encoded connection changes the output and restoring it reverses the result. On MNIST, the integer model achieves 97.16% accuracy across 10,000 test digits; Doomformer’s computation matches that model exactly on every value checked, which is distinct from whether the classification itself is correct. The project reports exact results for thousands of runs and nearly two million compared values in edge-case testing, as well as tests that deliberately alter weights or corrupt readouts to ensure errors are detected. These results demonstrate a carefully validated, inventive form of in-game computation, not a practical alternative to conventional neural-network execution.