NeuraMesh separates the market read path from the transaction and execution paths. This avoids treating an indexer snapshot as proof of a completed GPU workload.

// 01

Market read path

An indexer would assemble searchable capacity pools from chain state and provider attestations. Its API is a convenience layer. A stale index must never authorize new issuance or falsely display remaining supply.

// 02

Claim write path

Sui Move objects would represent source pools, capacity claims, and reservations. Invariants should prevent issued plus reserved usage from exceeding certified deliverable capacity; transaction builders must name the intended objects explicitly.

// 03

Execution plane

A scheduler maps a redeemed claim to an isolated worker and delivers the workload reference. Agent identity, placement, timeout, retries, and failure accounting belong off-chain, not in browser-side marketplace JavaScript.

// 04

Evidence and settlement

Walrus can hold referenced workload and receipt data while Sui records claim state. An independent metering and challenge path is necessary before settlements can fairly resolve disputes.

// Straight answers

Questions, answered.

Does Sui run model inference?

No. Sui is proposed for claim ownership and settlement; GPU workloads run on provider infrastructure.

Does storing a receipt prove delivery?

No. A stored receipt is evidence to inspect, not by itself independent proof that contracted compute was delivered.

Primary sources

External sources describe their own products; NeuraMesh implementation status is separate.

Explore the reference market

Compare the contract, not just the GPU.

These pages explain a proposed infrastructure design. Browse the reference marketplace without assuming a listing is available to buy.

Explore marketplace →