AI Layer

KAURAX includes an optional application layer for AI services. It is application contracts on an EVM chain, not a protocol feature, and it is not required by anything else in the system.

What belongs on chain, and what does not

Inference does not happen on chain. It cannot, at any useful scale, and pretending otherwise would be dishonest. What a chain is genuinely good at is the surrounding coordination:

On chain	Off chain
Identity — who is this provider or agent	Model weights
Authorization — who may call what	Inference execution
Payment and settlement	Prompt and response content
Reputation metadata — attestations, counts	Result verification
Service discovery	Serving infrastructure

AI provider
    │  registers a service, sets a price
    ▼
KAURAX  ── AIServiceRegistry ── AIAgentRegistry
    │            │
    │            ▼
    │       payment escrow / release
    ▼
application  ── consumes the service off chain, settles on chain

Contracts

blockchain/contracts/src/ai/:

Contract	Responsibility
`AIServiceRegistry`	Providers register services: metadata URI, price per call, active flag
`AIAgentRegistry`	Agents register an identity and an owner, and are authorized to spend
`AIPayments`	Escrowed, per-job payment: fund → deliver → release or refund

Honest limits

No oracle, no verification. The chain cannot tell whether a provider actually served

a request. AIPayments escrows funds and releases them on the consumer's acknowledgement or a timeout; it does not adjudicate quality or delivery. - Reputation is metadata, not proof. Counters and URIs, recorded on chain. Nothing attests that the underlying work happened. - No inference, no model hosting, no data storage. KAURAX stores pointers and money. - Unaudited, and unused. No AI service is deployed on any KAURAX network. The registry is empty, and the explorer will say so.

Why an L3 for this

Coordination workloads are high-frequency and low-value-per-call: an agent settling thousands of small payments needs blockspace that is cheap and not contended by unrelated applications. That is the case for an application-specific L3 generally, and it applies here as well as anywhere.

It is a reason to build on an L3. It is not evidence that anyone has.