Start here

↑ ↓ to choose · Enter to open · Esc to close

Trellis whitepaper

How a model request becomes a USDC payment and a verifiable record of the call.

Abstract

Two calls to the same model can travel through different inference services. A caller reconciling an expense needs the price and the provider. A caller sending confidential material also needs to know who can read the request.

Trellis exposes those decisions through one API. The catalog describes the available routes, the quote sets the terms of payment, and the signed receipt records the outcome.

From model to endpoint

A model ID names the requested model. An endpoint is a place that can serve it. Clients read the catalog for supported inputs, context length, pricing and available routes. A catalog entry describes an offer at the time of lookup. Endpoint availability can change before dispatch.

The router filters candidates using the request requirements and available supply, then considers signals including price and uptime. A client can set a minimum routing lane to express the conditions under which it will send a prompt.

A quote for one call

An unpaid chat request can return HTTP 402 with machine-readable payment terms. An x402 client checks the network, asset, amount, recipient and requested resource against its own spending policy. The payer authorizes the transaction and the client presents the signed payment with the matching request.

For example, a research tool can cap the output length and ask for a quote on one short summary. The tool can decline the quote without paying. If the user approves it, the tool can retain the payment response and receipt ID together so a later reviewer can reconcile the wallet activity with the model call.

Solana transfers expose transaction details, including account addresses and amounts. A public wallet address can link payments over time. Keep wallet signing credentials inside the wallet or signing service.

What a receipt establishes

The router signs the receipt with Ed25519. A verifier checks the signed bytes against the public key identified by the receipt. A changed claim fails the signature check. The signature establishes that the holder of that key made the claim. Assessing the answer’s accuracy or proving which hardware ran the model requires separate evidence.

Trellis groups receipt hashes into a Merkle tree. An inclusion proof combines a receipt’s hash with sibling hashes to reconstruct the batch root. At launch, the anchor process publishes roots in Solana memos. Checking final publication requires verifying the transaction, its signer and the root on the expected network. A newly issued receipt may precede its batch anchor.

Check a receipt

Disclosure and retention

Transport encryption protects data in transit. The systems performing public-route inference can read the request. Model and route disclosures, the router’s accounting records and the inference service’s own retention practices each answer a different privacy question.

Recovery also matters. A lost connection can leave the caller with a settled payment and no answer. Eligible non-streamed responses can be retained temporarily for wallet-authorized recovery. Review recovery limits before deciding what to send, and treat receipt IDs and payment references as shareable metadata. Read the data inventory.

Launch scope

The November 2026 launch supports public-route inference with per-call USDC payments. Check live model supply before sending a request with particular routing requirements.

See planned capabilities