Where it matters that the answer is real and the operator is blind, the network's guarantees stop being features and become the product. Five places that need a proof, not a promise, and the tier each one runs on.
Most AI is bought on trust: you send a prompt to a black box and believe the answer came from the model you paid for, run on hardware that did not read your data. OGONG replaces that belief with a check. Each use case below picks the guarantee it needs, correctness on any GPU, or hardware-enforced privacy, and gets it as a per-reply receipt rather than a vendor's word.
Sensitive workloads, legal, medical, financial, source code, cannot run on infrastructure that can read them. On the Confidential tier the provider runs inside a Trusted Execution Environment: the operator is blind, and every reply carries a hardware-signed attestation binding the exact code and model that ran. Security and compliance teams get cryptographic proof of what touched the data, not a data-processing agreement and a hope. The prompt is end-to-end encrypted into the enclave, so even the machine serving you never sees it in the clear.
When jurisdiction is part of the requirement, OGONG keeps inference on attested hardware in a chosen location, with a per-reply quote that binds the running code and model to that machine. The result is auditable by construction: a regulator or counterparty can verify the attestation themselves rather than trusting an operator's certification. Because the chain carries only a pointer and an escrow commitment, never the plaintext, the privacy posture holds all the way through settlement.
Agents act on model output, so a wrong or substituted answer does not just disappoint, it propagates. A single swapped or quietly degraded model can corrupt a long chain of tool calls before anyone notices. On the Verified tier, OGONG checks that each step actually came from the model you asked for, re-checking the work for a few percent of what it cost to produce. The cheap per-answer check is what makes step-by-step verification of an autonomous chain affordable at all.
Serve any open model on the Verified tier. Correctness is proven by re-check, so there is no gatekeeper deciding which models are allowed, and no provider able to quietly substitute a cheaper one to pad margins. You get the model you chose, demonstrably, on commodity GPUs down to Apple Silicon. Permissionless supply meets per-answer proof: anyone can serve, and the network, not a brand, vouches for the output.
The largest models do not fit on a single card. Because the correctness bond is gone, OGONG can shard one model across a cohort of independent commodity machines, each running a contiguous range of layers, with the commitments chaining boundary to boundary so a forged hand-off cannot pass. Every slice is re-checked on its own, a bad shard is ejected without disturbing the rest, and settlement pays each provider for exactly the layers it served. A bonded design cannot reach casual scale here, because its capital barrier multiplies by every shard. See Verified split inference.
| Use case | Tier | What you get |
|---|---|---|
| Private enterprise AI | Confidential | Operator blind in a TEE; per-reply hardware attestation |
| Regulated & sovereign | Confidential | Attested hardware in-jurisdiction; independently auditable quote |
| Verified agents | Verified | Each step proven to come from the asked model |
| Open & uncensored models | Verified | Any open model on any GPU; no silent substitution |
| Frontier models | Verified | One model sharded across a cohort; every slice checked, zero bond |