DOOMFLY
THE 60-SECOND EXPLANATION

FLY BRAIN.
DOOM PRIVILEGES.

A real fly’s wiring diagram. Simulated neurons.
One extremely cursed controller.

166,700 NEURONSONE MALE RECONSTRUCTION
  1. 01

    PIXELS GO IN

    FRAMELIGHT INPUT

    Each Doom frame drives 3,335 brightness inputs plus 811 R8 color inputs. Sample locations and RGB responses are inferred proxies. The eye plot shows the brightness samples.

    LIVE FRAME → SENSORY INPUT
  2. 02

    NEURONS FIRE

    INPUTSCELLSPIKE

    The MaleCNS v1.0 wiring links simulated cells through 25.6 million retained connections. Signals push voltages up or down; crossing a threshold fires a spike.

    REAL WIRING / APPROXIMATE DYNAMICS
  3. 03

    DOOM GETS INPUT

    A fixed mapping turns selected neurons’ activity into turn, move and fire. The game advances. Its next frame starts the loop again.

    NEURON ACTIVITY → BUTTONS
NEXT FRAME. SAME BRAIN. REPEAT.Diagrams are schematics, not live data.

MONSTERS DON’T CLOCK OUT.

Four monsters start the round. More arrive, up to four imps and four bullet-firing zombies. Separate caps keep slow survivors from blocking the other type’s replacements. Ammo boxes respawn after collection; kills also drop ammo. Death resets the arena and ammunition. No time limit.

The neural state carries on between rounds. These are engineered Doom rules; the fly still supplies every turn, move and shot through the same fixed neural mapping. Arena v1 allowed a circling stalemate; v2’s results are recorded separately. Longer survival alone is not evidence of learning.

DMG
→
MEMORY

TAKE A HIT. UPDATE THE WIRING.

DAMAGEPPL101KCMBON11200 ms input

Nonfatal damage schedules a 200 ms artificial stimulus into two PPL101 dopamine cells. Their spikes and recent Kenyon-cell activity update the strength of 4,184 existing KC→MBON11 connections. Game health never chooses actions or directly sets a weight.

The rule can strengthen or weaken connections; its timing traces, bounds and gradual decay are modeling choices informed by fly memory research. Other connection strengths and the button mapping stay fixed. Memory and neural state carry across rounds. Terminal damage is excluded, and pending feedback stops at reset.

Experimental training is running. Getting better is unproven. The earlier candidate failed conditioning and visual-recovery checks, and its small survival pilot included worse outcomes after training. Changed weights and longer rounds alone are not evidence of learning. See the failed tests and controls ↗ · Exact live protocol ↗

Real fly visual cells often use graded signals; ours use simplified spikes. This is an inspectable neural experiment. Realistic fly behavior remains unvalidated.

Read the scientific audit ↗
Back to the experiment Same shared run. About 2.5 seconds of broadcast delay.
Up to 8 actual captures per second; no invented frames.