PATENT PENDING · 9 ASPECTS

Technology
Stack.

Six cryptographic pillars. Privacy for model weights. Privacy for user inputs. Integrity proofs. Verified routing. All running on consumer hardware. No trusted hardware required.

SYSTEM ARCHITECTURE OVERVIEW


  USER / AI AGENT                     IOCHAIN NETWORK                      BLOCKCHAIN
  ┌─────────────┐                   ┌─────────────────────────────┐       ┌──────────────┐
  │             │  FHE(intent)  ►   │  SOLVER NODES               │       │              │
  │  enc(input) │  ─────────────►   │  dot(enc(intent), caps)     │       │  Model       │
  │             │                   │  → enc(scores)              │       │  Registry    │
  └──────┬──────┘                   └─────────────┬───────────────┘       │              │
         │                                        │ route to best node     │  Revenue     │
         │  FHE(input)                            │                        │  Distribution│
         └───────────────────────────────────────►│                        │              │
                                        ┌─────────▼───────────┐           │  ZK          │
                                        │  COMPUTE NODE       │           │  Commitments │
                                        │  ┌───────────────┐  │  Merkle   │              │
                                        │  │  WASM Sandbox │  │  root ►   └──────────────┘
                                        │  │  Layer 01/N   │  │
                                        │  │  Layer 02/N   │  │  Sub-proof distribution
                                        │  │  ...          │  │  ─────────────────────►
                                        │  │  Layer N/N    │  │           ┌──────────────┐
                                        │  └───────────────┘  │           │ PROVING NODES│
                                        │                     │   ◄────── │ ZK sub-proof │
                                        │  enc(output)        │           │ per layer    │
                                        └─────────────────────┘           └──────────────┘
                                                 │
                           enc(result) ◄──────────┘
P1EXECUTION LAYERFirst Aspect

WebAssembly Inference Sandbox

Models execute inside an isolated WASM linear memory space with hardware-enforced boundaries.

The Problem

Running untrusted model code on consumer hardware is a security nightmare. You cannot ask users to install arbitrary binaries from unknown developers. And model owners cannot trust that deployers won't extract their model weights.

The Solution

Neural networks are compiled into WebAssembly binary modules. WASM provides a deterministic, sandboxed execution environment with an isolated linear memory space — enforced by the runtime, not by OS permissions or hardware TEEs. The model runs as a WASM module, and its memory is strictly bounded.

Architecture Obfuscation

The compilation process applies a dependency-preserving permutation to the computation graph, inserts dummy nodes, and embeds a weight index mapping table in obfuscated form. Even if an adversary extracts the WASM binary, they cannot determine the true model architecture or separate genuine weights from dummy values.

TECHNICAL PARAMETERS

Execution envWebAssembly (WASM) runtime
Memory isolationLinear memory, hardware-enforced
ArchitectureObfuscated at compile time
Dummy nodesInserted probabilistically
Deployment targetBrowser, server, edge

COMPETITIVE LANDSCAPE

CAPABILITYIOCHAINCloud (AWS/GCP)Traditional DePIN
Consumer hardware support✓✗~
Browser-native execution✓✗✗
Model weight privacy✓✗✗
User input privacy (FHE)✓✗✗
ZK integrity proofs✓✗~
Private semantic routing✓✗✗
Per-request micro-payments✓✗~
CDN-like geographic distribution✓✓ (expensive)~
No installation required✓✗✗