Zero-Trust Attestation
TEE
Verification
Independent cryptographic verification of hardware attestation evidence from confidential inference providers. Raw quotes are parsed and verified against vendor roots of trust — no server-side claims are trusted.
Verification Methodology
Tinfoil Enclaves
Fetches .well-known/tinfoil-attestation from each enclave. The response contains a gzipped, base64-encoded hardware quote. For TDX enclaves, the quote is verified through Intel DCAP Quote Verification Library (dcap-qvl) using Intel PCS collateral (override via PCCS_URL if needed). For SNP enclaves, the report is verified against AMD KDS VCEK certificates with builtin ARK/ASK roots. TLS certificate SPKI fingerprints are compared to the attested report_data.
Venice.ai / Redpill / NEAR AI (dstack)
All three providers use the same dstack attestation format. Venice returns a direct attestation; Redpill and NEAR AI return a two-layer format (gateway + model attestations). In all cases, the raw intel_quote hex is parsed as a TDX quote and verified via dcap-qvl using Intel PCS collateral (or PCCS_URL override). The quote-bound tcb_info.app_compose hash is recomputed and compared to tcb_info.compose_hash; when compose-manager actions_hash evidence is present, that is recomputed too. Provider server_verification and verified fields are completely ignored.
NVIDIA GPU Attestation
GPU evidence payloads are submitted to NVIDIA's NRAS (NVRemote Attestation Service) which returns a signed JWT. The JWT signature is verified against NRAS's JWKS, the issuer is pinned to nras.attestation.nvidia.com, and the eat_nonce is checked against the request nonce. NRAS currently rejects dstack/Phala GPU evidence (unsupported architecture), so GPU verification is reported separately from TDX.
Chutes / Maple / Privatemode / Other Providers
Chutes is checked by fetching a live TEE chute's evidence endpoint, verifying the Intel TDX quote with dcap-qvl, checking that report_data binds the verifier nonce plus the ML-KEM E2E public key, and verifying NVIDIA GPU evidence when present. Chutes checks stay partial unless the workload image is digest-pinned or equivalent signed provenance is exposed. Maple Nitro checks also stay partial until expected PCR/workload provenance is configured. Privatemode is checked with a pinned Contrast CLI: the Coordinator is remotely attested, the active manifest must match the CDN manifest hash, and digest-pinned images are extracted from the verified initdata before the local proxy model path is accepted. PPQ is not counted as verified from liveness alone.