Neural Chain Architecture
Three-layer blockchain architecture powering decentralized AI inference
Three-Layer Architecture
AI Inference Layer
PyTorch + ONNX Runtime
- •Model registry and version control
- •TieredModelLoader (TorchScript, ONNX, llama.cpp)
- •Performance prediction and routing
- •KV-cache reuse and batching
Privacy & Consensus Layer
TEE + Zero-Knowledge Proofs
- •RSA 2048-bit + AES-256 encryption
- •Zero-knowledge proof verification
- •BFT consensus (80% threshold)
- •Node tier system (Gold/Silver/Bronze)
Universal Mining Layer
Dynamic Proof-of-Inference
- •Dynamic reward calculation
- •Performance-based multipliers
- •VRF-based random selection
- •Three-tier mining architecture
Core Components
Storage Network
IPFS content-addressed storage with caching, manifest support, and checksum management
Consensus Protocol
Byzantine fault-tolerant consensus requiring 80% agreement (8/10 miners, 40/50 verifiers)
Secure Enclaves
TEE attestation and threshold key management for protecting model secrets
Query Lifecycle
Client SDK → Network selection → Miner inference → Proof verification → Reward distribution
Query Data Flow
Query Submission
Client uploads encrypted payload off-chain and submits query hash on-chain
Miner Selection
Network selects eligible miner based on staking and infrastructure score
Inference Execution
Miner runs inference in TEE or secure software environment
Proof Generation
Miner generates zero-knowledge proof and posts it on-chain
Verification
Verifier validates proof and attestation deterministically
Finalization
On-chain module finalizes query, mints rewards, or initiates slashing