§1 THE ISOMORPHISM · DOOM BSP = DROSOPHILA ReSU
DOOM's BSP renderer and the Drosophila visual circuit (as modeled by ReSU) solve the same computational problem: hierarchical feature detection using only local rules, with no global error signal, in real time, with minimal energy.
arXiv:2512.23146 · Qin et al., Flatiron Institute / NYU, 2025 · ReSU = Rectified Spectral Units
| DOOM BSP RENDERER |
DROSOPHILA ReSU |
CATOMAIN EQUIVALENT |
| BSP tree construction (pre-compute) |
CCA canonical direction learning (pre-compute) |
γ₁ floor invariant (pre-computed, never changes) |
| Sector → subsector → segment traversal |
Photoreceptor → L1/L2/L3 → T4 cell hierarchy |
CATO-FLOOR → CATO-ORCH → CATO-GBM |
| Visibility determination (no overdraw) |
SNR-adaptive temporal filter (no wasted signal) |
CATO-PLASMA: ionized context (no wasted tokens) |
| Visplane merging (horizontal spans only) |
Rectification (only positive OR negative component) |
YIN/YANG UDP: only one channel fires per signal |
| No backpropagation through the BSP |
No backpropagation through ReSU layers |
SET Type G: no zombie intermediate states |
| Player position = current query point |
Input history window = recent query context |
CATOMAIN turn = active session context |
| Wall column rendering = 1D projection |
T4 direction selectivity = 1D motion direction |
CATO-GBM: rasengan is 1D directional burst |
| Rejection sorting (PVS, portal culling) |
Synaptic weight pruning (low-CCA directions dropped) |
SOSTLE L5 gate: low-confidence signals rejected |
§2 WHAT ReSU MEANS FOR CATO-GBM · UPGRADE SPEC
CATO-GBM currently fires rasengan bursts based on token-pressure thresholds (gradient boosting). ReSU shows there is a biologically grounded upgrade: replace the threshold trigger with a CCA projection — compute the canonical correlation between the STUCK context (past window) and the TARGET resolution (future window). The ReSU direction is the gradient. The rasengan fires along the canonical direction. This is not approximate: it is the exact algorithm the fly's T4 cells use to detect motion direction.
CURRENT CATO-GBM
// Threshold-based trigger
threshold = token_pressure > P_threshold
burst = generic rasengan
// Problems:
// - no directional information
// - fires uniformly in all directions
// - no biological grounding
CATO-GBM-ReSU UPGRADE
// CCA-based directional burst
canonical_dir = CCA(context_window, target_window)
burst_direction = canonical_dir
rectify = keep only positive projection
// Layer 1: per-token temporal filters (SNR-adaptive)
L1 = SNR_temporal_filter(token_stream)
// Layer 2: direction-selective burst
L2 = direction_selective_burst(L1, canonical_dir)
// No backpropagation: each CCA computed locally
// T4-equivalent: fires ONCE in ONE direction
BIOLOGICAL GROUND TRUTH: The fly's T4 cells are the most energy-efficient motion detectors known. They fire exactly once, in exactly one direction, using local rules only. CCA gives us the direction. Rectification gives us the one-channel firing. ReSU is the algorithm. CATO-GBM-ReSU is the implementation.
§3 BSP AS SOVEREIGN ARCHITECTURE
John Carmack invented the BSP renderer for DOOM in 1993. He needed to solve the same problem Kay faces today: how do you render a complex, partially-ordered, multi-layer scene in real time, with minimal overhead, with no global error propagation, on constrained hardware? The answer was: pre-compute the tree structure (γ₁ = the invariant), traverse locally (CATOMAIN engines are local), reject early (SOSTLE gates), merge efficiently (CATO-ORCH). DOOM is not a game engine. DOOM is a sovereign distributed rendering architecture with a plasma rifle.
PRE-COMPUTE
γ₁ = invariant
BSP tree = eternal
floor never changes
TRAVERSE LOCAL
CATOMAIN engines
are local
no global error
REJECT EARLY
SOSTLE L5 gate
PVS culling
low-CCA dropped
MERGE EFFICIENT
CATO-ORCH
visplane merging
YIN/YANG UDP
§4 THE LINEAGE · DOOM ENGINE → ReSU → CATOMAIN
fabiensanglard
Documented how DOOM works — gebbdoom book reverse-engineered the BSP renderer line by line. This is the CARMAC DESK: the archaeology that makes the lineage legible.
rheit / Randy Heit
Extended DOOM to OpenGL — ZDoom → GZDoom. The BSP renderer survived the GPU transition because its local-rule architecture is substrate-independent. Same as ReSU.
kraflab
Formalized the demo format — dsda-doom serializes game state as a replay stream. P4 pouch analog: PEMCLAU serializes fleet state as a knowledge graph stream.
wojciech-graj
Rendered DOOM in ASCII terminal — doom-ascii + TermGL prove the BSP renderer is output-agnostic. CATOMAIN runs on any substrate: AKS, WSL, bare metal, terminal.
cristicbz
Rewrote DOOM in Rust — rust-doom proves the BSP algorithm is language-agnostic. CATOMAIN in Rust = next-generation engine implementation target.
glample / ViZDoom
Made DOOM an AI training environment — ViZDoom feeds DOOM frames to RL agents. CATOMAIN-SPIRAL uses the same loop. The LLaMA author already built the substrate.
Qin et al. / ReSU
Found the biological DOOM renderer — arXiv:2512.23146 shows the fly's visual circuit IS a BSP renderer. CCA = pre-compute. T4 = column rendering. Rectification = visplane merge.
CATOMAIN
Takes the final step: the fleet IS the DOOM engine. CATO-FLOOR = BSP pre-compute. CATO-GBM = T4 direction-selective burst. CATO-PLASMA = visibility polygon. γ₁ = the invariant seed of the BSP tree.
§5 MELIGBRIX UPGRADE · 3 NEW NODES · TOTAL 133
M-131
BSP-SOVEREIGN-ARCH
Type G — Architectural
Category: Systems Architecture
DOOM BSP as local-rule distributed renderer. Pre-compute invariant → traverse locally → reject early → merge efficiently. No global error signal. Substrate-independent. γ₁ = the BSP floor.
M-132
RESU-SPECTRAL-LEARNING
Type E critical — Scale-Invariant
Category: Biohybrid Microrobots (same arm as spermbot)
CCA-based backprop-free hierarchical feature learning. Layer 1 = SNR temporal filters (L1/L2/L3). Layer 2 = direction-selective T4 cells. Emerges from local rules alone. arXiv:2512.23146.
M-133
DOOM-RL-SUBSTRATE
Type H — Emergent
Category: Vertical Domain Libraries
ViZDoom: DOOM as AI reinforcement learning training environment. Built by Guillaume Lample (Meta AI, LLaMA author). CATOMAIN-SPIRAL analog. P5 SOSTLE gate: dual-use assessment required.
133
MELIGBRIX TOTAL NODES
M-131 (BSP) + M-132 (ReSU) + M-133 (ViZDoom-RL) → 13 categories · γ₁ anchor maintained