Whitepaper

IO — A Privacy-First Multi-Model AI Inference Network

v0.1.0 · June 2026

Abstract

IO is a privacy-first AI provider with a single guarantee: nothing about a request is written down. It is reachable as an inference layer — from a web client, a CLI, a 21-tool MCP server, language SDKs and messaging relays — with no database, no logs, no analytics, no IP retention, and no accounts.

Every response carries a signed cryptographic receipt, io_receipt_v1, that attests which model answered, which policy was active, and that no prompt was retained. IO is not a single-model chat wrapper — it is a multi-model orchestration engine.

The network token, $IO, launches as a fair launch on PumpFun with no pre-sale and no venture allocation. This paper describes the design, the threat model it does and does not address, the verification mechanism, and the path from today's inspectable guarantees to tomorrow's hardware-attested ones.

Contents

1.Introduction
2.Design Goals and Non-Goals
3.Products — Chat, Code, MCP, and planned surfaces
4.Provider Capabilities and $IO Utility
5.The Model Lane — Gemini Image, GPT Image, Claude, and more
6.Architecture — memory-only, Rust, vLLM, Blackwell GPUs
7.Cryptographic Verification — io_receipt_v1, Ed25519, TEE roadmap
8.The $IO Token — fair launch, payment, access, burn, governance
9.Doctrine — six principles
10.Comparison — IO vs baseline
11.Roadmap — v0.1.0 through v1.5.0