Prompts arrive over AntSeed's end-to-end encrypted transport (X25519 + AES-256-GCM) and are proxied in memory. Never written to disk, logs, or databases. We persist only token counts, model name, latency, and a pseudonymous peer id.

Inference runs exclusively on Venice AI private-tier capacity — a provider contractually committed to not logging prompts. Upstream credentials live in memory-only pools; after the TEE migration they will exist only inside the enclave.

Responses stream straight back through the encrypted channel. Nothing is cached or stored — no response bodies, no embeddings of them, no "abuse-prevention" retention.

Content is never persisted. Operational metadata only: token counts, latency, error codes. Verifiable after TEE migration via remote attestation of the exact running code.

The TEE operator image will ship distroless: no shell, no SSH. Even we won't be able to inspect a running request. Until then, access is single-operator, key-only, and audited.

API users fund their own buyer address in the AntSeed escrow contract on Base — we never custody deposits. Your buyer key backup lets you withdraw without our involvement, even if we vanish.

The request sanitizer and the gateway source publish with the TEE launch — read exactly what is kept and what is discarded, line by line.