Verifiable is not the same as private. Every model carries a tier that says exactly what you are getting. There is no just-trust-me tier.
Every model on OGONG carries a trust tier that states precisely which guarantee the consumer obtains. The tier characterizes how the work is served and verified, not which organization produced the model. Two tiers exist, one for each method of establishing correctness, and there is no "just trust me" tier.
| Tier | Privacy | Correctness | Hardware |
|---|---|---|---|
| Confidential (TEE) | Operator is blind (enclave) | Hardware attestation | TDX / SEV-SNP + NVIDIA CC |
| Verified | Operator sees the content | Statistical audit (re-check) | Any GPU, incl. Apple Silicon |
The provider runs inside a Trusted Execution Environment and produces a hardware-signed attestation that proves what code and model are running and gives you an encrypted channel into the enclave. The operator cannot read your prompt or the response, and the enclave signs a per-reply receipt, so correctness is hardware-attested for the specific answer you got. Use it for anything sensitive. See Confidential computing.
The public-compute tier, on any GPU and no special hardware. The provider commits to its work and a verifier re-checks it cheaply by teacher-forcing a single pass. This proves the answer came from the claimed model. The guarantee is probabilistic, not cryptographic, and the serving machine sees the content; your identity is stripped at the router, but the prompt text is visible to the GPU running it. If you need privacy, use Confidential.
Because you cannot have it. Hiding content from the operator requires a TEE. On an ordinary GPU the operator can always read what the card is processing, so the Verified tier can prove correctness but never privacy. We are honest about this rather than blurring it. The stated endgame is zero-knowledge inference, which would give cryptographic correctness without a TEE, but it is not practical at language-model scale yet.