DOOMFLY

THE CONNECTION

Live game pixels → inferred retinal inputs → the full retained MaleCNS v1.0 graph → fixed game controls. 25,582,938 connections; 124,177,617 synaptic contacts.

DNp20 difference turns. DNpe017 rate moves; a spike fires. These four cells were chosen after visual-response calibration. Biological walking and mouthpart readouts remain visible as a comparison; they were silent in the initial test. No enemy coordinates, navigation policy or additional aiming policy enters the decoder. Native game weapon rules remain in Doom.

THE APPROXIMATION

This is reconstructed wiring with approximate neuron dynamics, not a validated fly emulation. Visual cells that use graded signals are modeled with a coarse spiking proxy. Retinal geometry, brightness response and tonic lamina current are assumed. The historical baseline produced controls with black input. The training candidate adds inferred R8 color input, KC adaptation and modeled chemical modulation; useful vision remains unvalidated.

42 unmapped photoreceptors remain in the graph. 3,718 transmitter signs are uncertain. Female BANC retinal mapping is unresolved; this run uses one male reconstruction.

REWARD / OFF

The baseline has fixed synapses. Stimulus delivery alone does not establish learning.

THE ARENA

The combat arena starts with four ranged monsters. Every three game seconds, another can spawn, up to four imps and four bullet-firing zombies. Each type has its own cap so slow survivors can’t block the other type’s replacements. Ammo boxes refill on contact and return eight seconds after collection; dead monsters also drop ammo. Death restarts the arena with 100 bullets. No round timer.

These are game rules, not fly behaviors. Neural activity continues across rounds. In experimental training, memory efficacies also persist; the other connection strengths stay fixed. Arena v2 fixes a circling stalemate in v1; their results stay separate. Inspect arena rules ↗

THE CLOCK

1× means simulation time matches wall time. Doom slows with the brain; no neural steps are skipped. All viewers watch one shared run and cannot control it.

The viewer buffers about 2.5 seconds of actual frames, capped at eight captures per wall second. Original capture timing is preserved; no frames or neural signals are invented. Enemy counts include monsters outside the camera view. The host must stay online.

Live configuration ↗