Cevell Technical Architecture

Cevell is an immutable, hardware-attested minimal confidential computing operating system designed specifically for private AI inference workloads on cloud and on-premises infrastructure.

1. Dual-Isolated Silicon Execution (CPU & NVIDIA GPU)

Cevell secures both CPU orchestration and GPU tensor calculation without trusting the host hypervisor:

  • Confidential CPU Instances: Intel TDX and AMD SEV-SNP enforce hardware memory encryption (AES-128/256-XTS) and page table integrity. Initial boot state is verified via Unified Kernel Image (UKI) measured into MRTD/RTMR registers or Launch Digest.
  • Confidential GPU Compute: NVIDIA Hopper (H100/H200) and Blackwell (B200) operate with hardware Confidential Protected Range (CPR) VRAM encryption and PCIe/NVLink bounce buffers. Device authenticity is validated via DMTF SPDM 1.2 and a 5-tier DICE certificate chain rooted in silicon fuses.
  • In-Memory Ephemeral Keys: The CVM generates an ephemeral X25519 keypair at boot strictly inside CPU-encrypted volatile RAM. Keys are never written to disk and are destroyed on shutdown.

2. Attestation Reports & 64-Byte REPORT_DATA Anti-Splicing Composite

Every CVM issues an In-toto Statement v1 attestation document via GET /attestation?nonce=<32-byte-hex> containing the hardware silicon quote, endorsement certificate chain, and NVIDIA SPDM evidence. Silicon registers enforce a strict 64-byte limit resolved via a dual-partition composite:

  • REPORT_DATA[0:32] = SHA-256(SubjectPublicKeyInfo(TLS_Cert)) — Physical TLS socket channel binding preventing MITM proxies.
  • REPORT_DATA[32:64] = SHA-256(HPKE_Pub || Nonce || GPU_Evidence_Digest || Auth_Pub) — Anti-splicing lock welding CPU, GPU, session nonce, and tenant authorization together.
  • Hardware Quote Offsets: USER_DATA is validated at byte offset 568..631 in Intel TDX Quotes and offset 80..143 in AMD SEV-SNP reports, cross-verified against official vendor Root CAs (Intel SGX/TDX, AMD KDS, NVIDIA) with live online CRL revocation checks.

3. Custom RFC 9180 Hybrid Public Key Encryption (HPKE)

All prompts and completions are sealed end-to-end using RFC 9180 HPKE (DHKEM(X25519, HKDF-SHA256) + AES-256-GCM / ChaCha20Poly1305):

  • Key Derivation: Derives req_key via info tag cevell-hpke-req-aes-gcm, and response keys (resp_key and resp_base_iv) bound to the client's request nonce.
  • Streaming Frame AEAD: Streaming tokens are framed as length-prefixed binary Protobuf messages. Frame nonces are derived by XORing the 4-byte sequence number into the base IV. Authenticated with a 13-byte AAD layout [seq][frame_type][timestamp_ns] enforcing strict monotonic sequence ordering to prevent replay or truncation.
  • Zeroization: All symmetric keys and intermediate cryptographic contexts are securely scrubbed from volatile memory on teardown.

4. OS Hardening: dm-verity, Zero-Console & Dynamic Hermetic Sealing

The rootfs is an immutable SquashFS filesystem verified in real-time block-by-block against a SHA-256 Merkle tree via dm-verity. Virtual serial consoles and SSH daemons are compiled out. Once model weights are loaded, invoking POST /v1/system/seal triggers atomic nftables rules dropping all outbound TCP/UDP egress.

5. cevell-tunnel Architecture & Transparent Local Proxy

cevell-tunnel provides an agnostic, payload-independent transport tunnel between local developer tooling and confidential CVM instances:

  • Local Proxy Daemon (cevell-tunnel proxy): Runs locally on 127.0.0.1:8080. Intercepts standard JSON HTTP requests from OpenAI SDKs, LangChain, Cursor IDE, Continue.dev, or cURL, transparently handling hardware attestation, HPKE Protobuf sealing, Ed25519 header signing, and streaming frame decryption.
  • Native SDKs: Python (cevell_tunnel.ConfidentialTransport) and TypeScript (@cevell/tunnel with createConfidentialFetch).
  • Wire Protocol: Zero-dependency binary Protobuf encoding for cevell.wire.v1.EncryptedInferenceRequest and StreamingInferenceFrame.