Native provider assignments for the accepted answering graph

Status: implementation in progress on development/provider-market. This is an opt-in research genesis extension, not an upgrade of the public 0.4.0 chain. TODO items 4–6 retain their original completion criteria.

The accepted graph currently executes on an operator-installed group of hosts. Clients must instead be able to discover signed, collateralized offers for the graph's logical owners, reserve a complete assignment and recover it without an SSH controller. Offers describe obligations, not proof of independent ownership or available hardware. The native ledger remains the assignment authority.

Assignment and recovery

Any funded key may register an HTTPS endpoint and certificate fingerprint, post collateral and advertise a rank of the currently accepted graph. Offers have a price, capacity and expiry. No operator allowlist grants provider membership. Clients may name explicit offers or ask native reservation to select available capacity atomically within their price ceiling. This is a bounded market choice, not an unbiased random committee or a claim of Sybil-resistant provider diversity.

A hosted inference reservation binds the existing expert job, complete graph, conversation, tokenizer, output bound, provider offers, prepaid complete audit and a bounded provider fee. Every assigned owner acknowledges the assignment before execution. Provider collateral is unavailable for other obligations while reserved. Identities, endpoints and certificate fingerprints are copied into the assignment: later advertisements cannot mutate an in-flight request.

Preparation and execution have native block deadlines. After either deadline, any participant can propose available replacement offers within the original provider-fee ceiling. The original request, graph and execution escrow stay fixed. The assignment epoch changes; receipts bind that epoch so delayed owners cannot settle the new attempt. Attempts and outstanding assignments are capped. Recovery restarts the deterministic request from its pinned model; it does not invent lost activations or accept a provider's cache as evidence.

Timeouts release assignments. Missing acknowledgement or a failed multi-owner execution is not by itself proof that an individual worker cheated. In particular, a client must not burn an honest provider's bond by submitting a request that fails the provider's numerical preflight. The coordinator's native claim bond is reserved with the assignment, then exposed to the existing adjudication rules when it submits an execution claim. Service fees pay only after accepted complete replay. No hosting transaction issues tokens, and every transition conserves supply.

Funding and delivery

The existing funded audit quorum must accept a complete budget before a hosted request reserves work. Its reservation remains attached to the job across a coordinator change; only the native replacement rule can change its publisher. A failed adjudicated attempt consumes its own audit obligation. Any retry needs another explicitly funded obligation, within the finite attempt limit.

Customer funding uses fund_hosted_audit: its budget commits to the payer, graph and complete request hash. Native reservation rejects another payer or request, and the generic audit lock rejects attaching that budget to unrelated work. The coordinator cannot consume a customer's verification reservation as a general training subsidy. An ambiguous rejected reservation is retired only after committed closure makes its budget permanently unusable; its original signed bytes and unknown transaction outcome remain recorded.

The existing token-based execution payment is supplemented by each provider's fixed fee for the complete bounded request: prompt processing, selection, auxiliary models, transfer and its stated availability period. A token count alone does not price all those costs. Retention beyond that service period, auditing and ledger fees must be separately specified and funded. Measured costs and a finite sponsorship policy are prerequisites for claiming item 5 complete.

The serving transport must authenticate the registered identities, enforce message and memory bounds, fence assignment epochs and restore only committed objects. Reusable model partitions may remain resident between requests; request state and conversation context remain separate. The client must distinguish provisional streamed delivery from final, replay-authorized ledger payment. Prompts and responses remain public under the existing native request format.

The implementation now uses signed, certificate-pinned TLS 1.3 frames. A provider's own synchronized full node authorizes its peers from one committed job/assignment snapshot; arbitrary remote RPC responses are not accepted as state proofs. Frames have explicit tensor/control limits, sequence numbers, assignment epochs, retry identity and a bounded mailbox. A stopped local chain or changed assignment stops authorization. After a process loses its frame state, execution requires a new native epoch. The durable transaction outbox preserves signed bytes when acknowledgement of submission is lost.

Five separate CPU processes now reproduce the fixed-group executor's complete outputs over this transport, including ordinary and multi-turn answering. Every owner signs the same graph/request/epoch/transcript statement. Fresh owners restore only their own model partitions from hash-addressed mirrors. These are numerical and protocol fixtures, not the remaining operated LLM failure/latency trial or an independent-operator soak.

The experimental runtime is described in the provider guide. The subsequent operated native preflight passes fresh-key replacement after coordinator and backbone-owner loss, followed by complete funded replay and exactly-once settlement on the CPU fixture.

Validation before an LLM allocation

• Native tests: open registration, complete assignments, collateral and supply, capacity exhaustion, graph pinning, prepaid audits, coordinator replacement, stale receipts, duplicate payment, failed audits, cancellation and expiry.
• Transport tests: authenticated bounded messages, wrong identity/epoch, corrupt objects, reconstruction on a replacement owner and coordinator restart.
• One committed serving trial reuses the accepted checkpoint, includes required backbone loss and concurrent clients, and freezes latency, availability and total-cost limits before allocation. It does not repeat learning or count a transport migration as a new learning cohort.

Independent administration cannot be demonstrated by additional keys or AWS instances under this operator. Item 4 stays open until independently administered operators complete its soak and ownership/voting-power requirements.

Relevant engineering precedents are Petals' fault-tolerant distributed inference and libp2p's binding of peer identity to secure channels. Neither supplies NeuroShard's settlement, verification or ownership guarantee.