Skip to main content
The Verification Center at cloud.near.ai/verify is a public web page that verifies NEAR AI Cloud attestation evidence for a model and, optionally, the signature on a single reply. It runs the same checks described in the other verification guides, so you can inspect the evidence without writing code or holding an API key.
The Verification Center is a convenience for inspection and spot checks. It does not replace the checks your application performs when it enforces its own verification policy.

Open the page

Add the model ID to the model query parameter:
URL-encode the model ID, so z-ai/glm-5.3-flash becomes z-ai%2Fglm-5.3-flash. Without a model parameter, and without a reply receipt that names a model, the page asks you to open it from an app that uses NEAR AI or to add the parameter. The page starts verifying as soon as it loads. Select Verify again to fetch new evidence and repeat every check.

How verification works

Verification runs in your browser with the NEAR AI inference SDK. Each visit fetches hardware evidence fresh rather than reusing a cached result, and the page shows details as trusted only after your browser has verified them. The page verifies two things: Model evidence is checked only for models whose attestation format the SDK supports. For other models the page verifies the Gateway and tells you that model verification is not yet supported.

Read the overall result

The heading at the top of the page summarizes everything the page checked. An evidence outage shows as Unable to verify and not as a failure, but a failed signature or model-binding check always takes priority over an outage.
Verified hardware evidence and signing keys do not establish that a deployment is an approved deployment of the named model. The Verification Center checks deployment measurements for consistency but does not compare them against a trusted allowlist. If you need that guarantee, apply your own measurement and provenance policy.

Hardware verification

The Hardware verification section has two tabs, Model Verification and Gateway Verification. Both show the same layout. The page opens on the model tab when model evidence verified and on the Gateway tab otherwise.

Summary

Signing identity

Signing address and Signing public key identify the key that this component uses to sign replies. These values are bound to the hardware quote, so a verified quote ties the key to the attested environment. Use them to match the signer of a response signature.

Evidence sections

Expand a section to see the raw evidence behind the verdict. Every value has a Copy button, so you can verify it with another tool.
The Intel TDX quote and the values it commits to: request nonce, report data, event log, OS image hash, compose hash, app compose, and any Intel security advisories. Links open Intel’s TDX documentation and the Intel TDX Explorer, where you can check the quote independently. See Quote, nonce, and signer bindings for how these values are checked.
Shown on the Model tab. NVIDIA’s Remote Attestation Service verified the GPU evidence, and your browser checked the signed result against NVIDIA’s public keys. The section lists the nonce, GPU architecture, evidence list, and raw payload, with links to NVIDIA’s attestation documentation and API.

Verify a reply

An application that uses NEAR AI can link to the Verification Center with a reply receipt, so that you can check one specific response in addition to the hardware evidence. The receipt travels in the URL fragment:
The fragment is base64url-encoded JSON with these fields: When a receipt is present, a Reply signature section appears above the hardware evidence. It runs three checks:
  1. Signature: your browser checks the signature over signedText. ECDSA signatures must be 65 bytes and recover the stated address. Ed25519 signatures are checked against the stated 32-byte public key.
  2. Model binding: the signed text must name the same model as the receipt and the verified deployment. A Gateway receipt whose signed text does not include a model can be only partially verified.
  3. Signer: the signing key must match a signer in the freshly verified evidence. A provider_tee signature must come from one of the verified model replicas, and a gateway signature must come from the verified Gateway. A Gateway signature on its own does not prove that the model ran in a TEE.
The section also lists the completion ID, signing address, request and response hashes (SHA-256), signed message, and signature, each with a Copy button.
The page checks the signature over the hashes in the receipt. It does not receive the chat content and cannot recompute those hashes, so only the application that sent the request can confirm the hashes match the original request and response. The verifiedAt timestamp is reported by the linking application and is not independently verified.If the signer cannot be matched, the replica that signed the reply may have restarted or rotated its key since. That does not by itself mean the signature is invalid. For a completion you want to verify in full, follow the response signature guide.
If the receipt is malformed, too large, or repeated, the page reports an invalid receipt and the result is Verification failed. Ask the application for a new verification link.

Privacy

  • The receipt is read from the URL fragment (#), which browsers do not send in page requests.
  • The page does not send the receipt to analytics.
  • Verification requests do not include chat content.

Troubleshooting

Next steps

Gateway attestation

Verify the gateway’s evidence in your own client.

Model attestations

Verify every model candidate before a completion.

Response signatures

Verify the signature on a completion.

Verification policy

Decide which evidence your application requires.