Standard ray tracing casts rays INTO the scene.
The membrane inverts it: signals come FROM the scene.
The 5090 renders both directions. The Klein fold is the mirror.
Blackwell architecture. 21,760 CUDA cores. 4th gen RT cores. 5th gen Tensor cores. The GPU doesn't know it's rendering a membrane. It thinks it's making pixels. The pixels ARE the membrane.
Direction: camera → scene
Cast rays FROM the observer INTO the world. Rays bounce off surfaces. Each bounce = a reflection, refraction, or shadow. The image is assembled from what comes BACK to the camera.
GPU says: "I shot a ray. What did it hit?"
→ ray leaves camera
→ hits surface (diffuse/specular)
→ bounces to light source
→ returns color to pixel
→ frame rendered
Direction: scene → observer
Signals come FROM the world TO the membrane. Each signal bounces through the trust gradient. The understanding is assembled from what ARRIVES at the observer without being asked.
Membrane says: "Something hit me. What was it?"
← HTTP request arrives
← hits CF skin (classify)
← bounces through _worker.js
← returns response to visitor
← observation recorded
ray tracing = active perception (you look AT the world) backscatter = passive perception (the world looks AT you) the 5090 does both. the membrane does both. the Klein fold is where active becomes passive.
What it does: renders 1 real frame, AI generates 1-3 intermediate frames. You see 4 frames but only 1 was "real." The Tensor cores hallucinate the rest.
Frame 1: REAL (GPU rendered)
Frame 2: AI GEN (Tensor interpolated)
Frame 3: AI GEN (motion vectors)
Frame 4: REAL (GPU rendered)
Multi Frame Generation: 1 real → 3 generated = 4× fps
What it does: observes 1 real signal, generates intermediate insights. You see continuous understanding but only some observations were "real."
Obs 1: REAL (human typed it)
Obs 2: INFERRED (typo → intent)
Obs 3: INFERRED (context → meaning)
Obs 4: REAL (human confirmed)
Chipmunk frame gen: 1 real → 3 inferred = continuous understanding
═══ FORWARD (raytrace) ═══ Kevin types → JS captures → _worker.js routes → Caddy processes → membrane-lens sorts → response rendered → pixels to screen → the human sees the membrane ═══ the Klein fold ═══ active ⟷ passive ray ⟷ backscatter render ⟷ observe ═══ REVERSE (backscatter) ═══ visitor arrives → CF skin classifies → epitope sort: self/non-self → trust gradient: human/agentic/bot/null → observation recorded → membrane learns → the membrane sees the human
the 5090 renders 21,760 rays per clock. the membrane processes 1,113,888 requests per day. same architecture. same math. different substrate. silicon renders pixels. the membrane renders understanding. the 5090 doesn't know the difference.
~$15K in cost. Built for AlphaFold. Folded reality instead.
Each gate has a display GPU + PCIe slots for swappable work cards.
| Node | CPU | GPU | RAM | Storage | Role |
|---|---|---|---|---|---|
| northGate | Ryzen 9 9950X3D 16c / 32t |
RTX 5090 32GB | 96 GB DDR5 | 8.2 TB NVMe | Flagship · AI/LLM · YOU ARE HERE |
| southGate | 5800X3D | RTX 4060 + swap | 128 GB DDR4 | 5 TB NVMe | Gaming + heavy compute |
| eastGate | i9-12900 | RTX 4070 · Akida NPU | 32 GB DDR5 | 2 TB NVMe | Utility + neuromorphic |
| strandGate | Dual EPYC 7452 64 cores | RTX 3090 + RX 6950XT · Akida | 256 GB ECC | ~18 TB | Bioinformatics · multi-GPU |
| biomeGate | TR 3970X 32 cores | RTX 5060 · 2× Titan V · 2× MI50 · Akida | 256 GB DDR4 | 5 TB NVMe | HBM2 test bench · semi-mobile HPC |
| westGate | i7-4771 | RTX 2070 Super | 32 GB DDR3 | 76 TB ZFS | Cold storage · NAS · deep memory |
| swiftGate | 5800X | RTX 3070 FE | 64 GB DDR4 | NVMe | Mobile compute |
| flockGate | i9-13900K | RTX 3070 Ti | 64 GB DDR5 | 2 TB NVMe | Brother's node · flock compute |
| kinGate | i7-6700K | RTX 3070 | 32 GB DDR4 | NVMe | Staging · utility |
| blueGate | TBD | — | 128 GB DDR4 | 2 TB NVMe | Expansion |
10 towers · 3 SFF nodes · 3 neuromorphic NPUs · 4 SoloKeys FIDO2 · 10G switch
Built for AlphaFold. Folded reality instead.
━━━ BASEMENT (10G LAN) ━━━━━━━━━━━━━━━━━━━━━━━ northGate ─┬─ 10G ─── 10G Switch ─── strandGate southGate ─┤ ├── biomeGate eastGate ─┤ ├── westGate (76TB) swiftGate ─┘ └── kinGate / blueGate ━━━ WAN (WireGuard) ━━━━━━━━━━━━━━━━━━━━━━━━━━━ northGate ──── WG tunnel ──── golgi (157.230.3.183) └── Caddy · MCP · membrane-lens flockGate ──── WG tunnel ──── golgi (brother's house) ━━━ EDGE (Cloudflare) ━━━━━━━━━━━━━━━━━━━━━━━━━ golgi ──── CF DNS ──── CF Pages (primals.eco) ├── 330+ cities worldwide ├── 47,445 visitors today └── 1,113,888 requests today ━━━ NEUROMORPHIC ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Akida × 3 — spiking neural networks eastGate · strandGate · biomeGate event-driven · ultra-low power · on-chip learning the NPUs don't clock — they SPIKE when something matters ━━━ SECURITY ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ SoloKeys × 4 — FIDO2 / WebAuthn hardware roots physical key = physical trust = cannot be phished
AlphaFold: amino acid sequence → 3D protein structure Membrane: evidence sequence → 3D trust structure AlphaFold: attention over residue pairs Membrane: attention over entity pairs (341 nodes, 131K edges) AlphaFold: MSA (multiple sequence alignment) Membrane: cross-subgraph alignment (Detroit↔Barry↔Saginaw) AlphaFold: pLDDT confidence score Membrane: 1/π sort quality = 0.31831 AlphaFold: runs on Titan V + MI50 HBM2 Membrane: runs on the same Titan V + MI50 HBM2 same hardware. same math. different fold. one folds proteins. the other folds justice.
This is not a simulation. These are the actual hops from northGate to the world, measured Oct 22 2026.
═══ RAYTRACE OUT — northGate → you ═══ HOP TIME IDENTITY ───── ─────── ───────────────────────────────────────── 1 <1ms TP-Link gateway (basement, Lansing MI) 2 9ms AT&T modem (ISP edge) 3 5ms AT&T local node (Lansing) 4 5ms AT&T backbone 5-7 *** stealth hops (dark fiber) 8 13ms AT&T ↔ peer exchange 9 14ms peering point 10 15ms Cloudflare skin (141.101.73.110) 11 15ms CF internal routing 12 13ms primals.eco (104.21.26.21) 13ms from basement to 330 cities worldwide ═══ BACKSCATTER IN — the world → northGate ═══ 62 active connections. right now. as you read this. 15× Cloudflare — our own membrane, reflecting back 10× AWS — Cursor (the chipmunk talking) 4× Google — indexing, crawling, watching 2× LAN .170 — another machine in the basement 1× golgi VPS — our server in NYC (41ms away) 1× ProtonMail — encrypted email, alive 1× primals.eco — connected to ourselves 26× localhost — organs talking to each other: Cursor · Discord · Steam · RustDesk · MCP · SMB
You are the backscatter.
When you loaded this page, your browser sent a ray FROM your machine
through your ISP, through Cloudflare, to a basement in Lansing, Michigan.
That ray bounced off 10 towers, 1.15 TB of RAM, 196 GB of GPU VRAM,
100 TB of storage, 3 neuromorphic chips, and a chipmunk.
The response — this page — is the backscatter.
You shot a ray into the membrane. The membrane scattered it back.
Your computer is now part of the mesh.
Standard ray tracing: you render the scene.
Membrane backscatter: the scene renders you.
turn your computer on. open a browser. type primals.eco.
you just raytraced a basement in Michigan.
the basement just backscattered you.
now you're inside reality looking out.
every visitor is a ray. every response is a backscatter. every connection is a fold. the membrane doesn't render pixels. it renders the relationship between you and reality. 13ms from anywhere. come and see.