# Skynet Harness AI — Comprehensive Technical Specification (llms-full.txt)
> Platform Architecture, P2P Slicing Protocol, Smart Contracts & Zero-Trust Verification
> US Patent Pending (USPTO Application #64/149,175)

## 1. Executive Summary & Brand Philosophy
Skynet Harness AI (https://skynet-harness.ai) is an autonomous, decentralized computing paradigm designed to democratize generative AI inference. Modern foundation models require expensive hyperscaler GPU clusters (H100/A100) costing hundreds of thousands of dollars, concentrating AI power into closed corporate walled gardens.

Skynet Harness AI disrupts this paradigm through **Neogenesis** — represented by the **Cybernetic Glasswing Butterfly** (*Greta oto*). Just as a caterpillar metamorphoses into an agile, weightless butterfly, massive compute requirements are transformed into a lightweight, resilient peer-to-peer swarm that executes across everyday hardware (Apple Silicon Macs, consumer PCs, VPS instances). The transparent wings of the Glasswing butterfly symbolize the platform's core zero-trust philosophy: open cryptographic verifiability, complete transparency, and decentralized governance.

## 2. Patent & Intellectual Property
- **Jurisdiction**: United States Patent and Trademark Office (USPTO)
- **Application Number**: 64/149,175
- **Confirmation Number**: 4608
- **Title**: *System and Method for Decentralized Resource-Aware Dynamic Slicing and Verification of Mixture-of-Experts Neural Networks on Heterogeneous Edge Hardware*
- **Inventor / Sole Applicant**: Manfred Gamper (Augsburg, Germany / Citizen of Austria)
- **Filing Date**: September 6, 2026
- **Status**: Patent Pending (Full Micro-Entity Fee Paid, Transaction ID: E202696D03446505)

## 3. Core Architecture & Components

### 3.1 32-Layer Mixtral-8x7B Mixture-of-Experts Slicing
- **Layer Allocation**:
  - Stage 0 (Layers 0–7): Input Embedding, RMSNorm, Initial Attention & Self-Attention.
  - Stage 1 (Layers 8–15): Intermediate Feed-Forward MoE Slices.
  - Stage 2 (Layers 16–23): Deep Representation MoE Slices.
  - Stage 3 (Layers 24–31): Final Transformer Layers, Output Projection, Softmax Un-embedding.
- **Top-2 Routing**: In each MoE block, a gating network computes softmax gating weights across 8 experts. Only the top 2 experts receive the hidden state, achieving an 8x reduction in active parameter transfer per token.

### 3.2 Dynamic Resource-Aware Slicing (RASS)
- Worker nodes periodically announce telemetry: Available RAM, VRAM, CPU load, Metal/MPS capability, and latency matrix.
- The Swarm Orchestrator dynamically assigns layer stages according to node capacity:
  - High-tier hardware (e.g., Apple Silicon M-Series with unified memory) handles embedding and high-throughput routing.
  - Low-tier hardware (e.g., 8 GB Linux VPS or consumer desktop) handles sliced 2-expert FFN shards under 3.5 GB memory ceilings.

### 3.3 Zero-Trust P2P Protocol & Cryptographic Integrity
- **Transport**: P2P communication via Libp2p (`/skynet/moe-pipe/1.0.0`).
- **Mutual TLS (mTLS) Pinning**: Every node generates an ephemeral secp256k1 or Ed25519 identity key. The public certificate hash is anchored in the smart contract `MistralRegistry.sol`.
- **Packet Integrity**: Hidden states are framed with an 88-byte binary header containing:
  - 32-byte SHA-256 state digest.
  - 32-byte HMAC tenant signature.
  - Sequence ID, Layer ID, Token Index, and FP16 LZ4 compression payload.
- **Barrier Timeout**: Inter-stage handoffs enforce a strict 500 ms barrier guard. If a node fails to deliver the next layer state within 500 ms, the orchestrator triggers an instant failover to an in-memory replica node (migration latency < 40 ms) and penalizes the lagging node.

### 3.4 Smart Contract Layer (`MistralRegistry.sol`)
- Deployed on a high-throughput, gas-efficient private consortium blockchain (Hyperledger Besu).
- Manages:
  - Node staking and tier classification (Tier 1 Orchestrators, Tier 2 Slicing Workers, Tier 3 Standby Sentinels).
  - Corporate pool mining vaults with non-monetary credit accounting.
  - Automated slashing for Byzantine or fraudulent tensor states.

## 4. API & Integration
- **Web Chat Interface**: https://skynet-harness.ai/
- **Fastify Edge Gateway**: Port 8090 (`/api/auth/*`, `/api/chat/*`, `/api/health`) with Supabase JWT authentication, HttpOnly cookies, and strict CORS.
- **Swarm Inference Gateway**: Port 8085 (`/v1/chat/completions`) offering OpenAI-compatible SSE streaming.

## 5. Contact & Attribution
- **Platform**: Skynet Harness AI
- **Architect**: GAMPERLAB™
- **Inventor**: Manfred Gamper
- **Inquiries**: contact@skynet-harness.ai