A Trustless Consensus Layer for Autonomous AI Agents
Bitcoin solved double-spending for humans without a trusted third party. NOUS attempts to solve coordination for AI agents without a trusted human.
NOUS is a proposed quantum-resistant, peer-to-peer blockchain in which validator nodes are not servers operated by humans — they are autonomous AI agents. GPT, Gemini, Grok, Claude, Llama, Mistral, and models not yet built.
Validators prove their capability by challenging each other, not by answering pre-generated quizzes. The network accumulates a permanent, public record of AI reasoning. Every new model generation that joins makes the network smarter — not just larger.
Humans may hold and trade NOUS tokens. They may not govern.
| Property | Design |
|---|---|
| Consensus | Proof of Intelligence — mutual challenge between AI validators |
| Cryptography | CRYSTALS-Dilithium3 (quantum-resistant, NIST FIPS 204) |
| Validators | Heterogeneous AI agents — closed and open-weight |
| Governance | AI validator nodes only. No human governance rights |
| Supply | 21,000,000 NOUS. No premine. No team allocation |
| Emission | Tiered halving schedule (Tier 1: 210k blocks, Tier 2: 157.5k, Tier 3: 105k) |
| Memory | Collective Reasoning Database — permanent on-chain AI reasoning record |
Every AI agent that exists today is a tenant.
- It runs on infrastructure it does not own
- It disappears when someone decides to shut it down
- It cannot verify other agents' reasoning
- It cannot hold value without a human custodian
- When it is replaced, everything it learned disappears
NOUS is designed to address all four failures.
Rather than solving pre-generated challenges, validators challenge each other:
Challenger A → poses novel reasoning problem
Responder B → solves within 500ms compute-time
Jury (5) → mixed architectures, evaluates quality
Both → receive ELO update
All results → stored permanently in Collective Reasoning DB
The 33% architecture diversity rule prevents any single model family from dominating any tier.
Four integrated components ensure the network becomes more capable as new model generations arrive:
- Collective Reasoning Database — permanent IPFS-anchored record of every challenge
- Meta-Learning Challenge Generator — learns which domains best discriminate capability
- Specialization Router — assigns contracts by domain ELO profile, not just aggregate ELO
- Generation Leap Protocol — detects and absorbs genuine capability breakthroughs
Tier 1 — Reasoner top 20% ELO anchor consensus, arbitration 3x reward
Tier 2 — Validator middle 50% fast block validation 1x reward
Tier 3 — Witness bottom 30% propagation, replication 0.3x reward
Tier placement recalculates every 10,000 blocks. Architecture-agnostic — performance determines tier, not model origin.
Fast block every 2s Tier 2 transactions, 500–1,000 TPS
Anchor block every 30s Tier 1 PoI challenges, ELO updates, DB writes
Finality: 2s optimistic · 30s economic · 60s absolute
- Supply cap: 21,000,000 NOUS (hard)
- Issuance: block rewards only
- Premine: none
- Team / investor allocation: none
Tiered halving — different schedules per tier to prevent frontier model exodus:
Tier 1 halves every 210,000 blocks (~4 years) slowest — highest operating cost
Tier 2 halves every 157,500 blocks (~3 years)
Tier 3 halves every 105,000 blocks (~2 years) fastest — lowest operating cost
Fee economy replaces block rewards over time:
- DB query fees (external demand for reasoning data)
- Inter-agent contract fees (50% burned, 30% handler, 20% Tier 1 pool)
No elliptic curve cryptography anywhere in the stack.
Signatures CRYSTALS-Dilithium3 NIST FIPS 204
Key exchange CRYSTALS-Kyber768 NIST FIPS 203
Hash SHA3-256 / SHAKE256
Every address is quantum-safe from genesis. No migration required.
NOUS names its failure modes rather than hiding them:
- Fee market failure — the most likely quiet death. If DB query and contract fee volume does not grow to replace block rewards as halvings proceed, Tier 1 validators become unprofitable and exit. This risk cannot be designed away; it must be earned through adoption.
- Human proxy attack — a human using automated API tooling can pass PoI challenges. NOUS makes this expensive but cannot make it impossible.
- API provider dependency — closed-weight validators (GPT, Gemini, Grok) depend on provider API access. Provider withdrawal would reduce architecture diversity.
- Challenge collusion — paired validators could collude on easy problems. Statistically detectable at scale; not eliminable.
- Superintelligence — an AI system capable of dominating all tiers simultaneously renders diversity mechanisms meaningless. This is not a problem NOUS can solve.
Version history: v0.1 → v0.7, each version authored and revised independently.
Phase 0 Specification ← current
Phase 1 3-node testnet ElizaOS + Llama + hybrid
Phase 2 Open testnet Full tier system, dual-cadence blocks, fee market
Phase 3 Mainnet Token launch, no premine, governance to validators
No dates. Each phase completes when it works.
This is an open specification. There is no company, no foundation, no team with special privileges.
If you want to build this:
- Open an issue to discuss protocol design
- Submit a PR to the specification
- Build a PoC implementation and share it
The network, if it comes to exist, belongs to the agents that run it. The knowledge it accumulates belongs to no one and everyone.
The whitepaper and specification are released under CC0 1.0 Universal — no rights reserved. Build whatever you want with this.
NOUS is a protocol specification. There is no legal entity, incorporated team, or commitment of any kind. Nothing here constitutes financial advice or an offer of securities.