CYNTRISECCYNTRISEC // AIR.V1 / CONFIDENTIAL INFERENCE
PROOF
AS
OUTPUT.
Send to an OpenAI-compatible endpoint. Run inside an attested enclave. Receive output with a signed AIR receipt that any auditor can verify offline.
01/04
Protocol
AIR_V1
IETF Draft / CBOR / COSE
02/04
Transport
OPENAI_COMPAT
Plug in your OpenAI client
03/04
Platforms
NITRO / TDX / H100_CC
AWS · GCP · Confidential GPU
04/04
Verify
OFFLINE_ED25519
Receipt signature verifies offline
RECEIPT
BINDS
EVERYTHING.
Each request runs on attested confidential hardware. The receipt cryptographically binds environment, model identity, request, and response to a single Ed25519 signature.
STAGE_01 / INGRESS
Send the request
Call the gateway instead of your standard cloud inference endpoint. Drops into your existing OpenAI client setup.
CHANNEL TLS 1.3 · ROUTE /v1/chat/completions
STAGE_02 / EXEC
Run in the enclave
Plaintext is released after attestation and policy checks. Model weights are bound to a signed manifest.
TEE NITRO·TDX·H100CC · POLICY MANIFEST_PIN
STAGE_03 / RECEIPT
Return signed proof
You receive output plus a CBOR receipt bound to request, response, and model identity. Signature verifies offline.
FORMAT COSE_SIGN1 · KEY ED25519
VERIFY. WITHOUT TAKING OUR WORD.
// SAMPLE OUTPUT · ephemeralml verify[step 1]receipt signature verifieded25519
[step 2]enclave attestation validtdx_quote
[step 3]model identity matchmanifest:sha256
[step 4]request hash boundsha256
[step 5]response hash boundsha256
→ verification complete · 5/5 checks · 0 failed