Verified replies

How every reply is checked against NEAR AI Cloud’s TEE attestation, what the badge levels mean, and how to open the proof on Phala’s TEE Explorer.

Every model call runs on NEAR AI Cloud inside a TEE. After a reply finishes, the engine fetches each completion’s signature and checks it against the hardware attestation (engine/src/verify.ts) before the app shows the Verified badge.

​

Badge levels

The badge shows the weakest step’s level, at the level the NEAR AI SDK proves:

Level	Meaning
model_tee	The model itself ran in a TEE and signed the reply
gateway_tee	NEAR AI’s gateway TEE signed the reply (streaming replies are checked at this level)

Only verifiable (Confidential or third-party TEE) models get a badge. Incognito models never do.

​

What is kept

The proof behind the badge is stored on the run’s run.verified event: signatures, hashes and Intel’s TDX quote. Never the conversation text. It is about 12 KB, mostly the quote, and it is never streamed: only the owner can read it, through GET /v1/runs/{runId}/proof.

​

Check it yourself

In the app: click Verified via NEAR AI TEE gateway under a reply, then View proof, then Check it in TEE Explorer. The backend uploads the Intel quote to Phala’s TEE Attestation Explorer (POST /v1/runs/{runId}/proof/tee-explorer) and the app opens the report. Only the quote leaves, no conversation text, and the report then shows in the Explorer’s public list of recent attestations.

A real report: proof.t16z.com/reports/1b374e5b… reads VERIFIED. Its report data ends with that run’s nonce (…99e266), and its MRCONFIG carries the compose hash (877959253d…) found in the proof.

​

Events

run.verified follows run.finished on the stream. See Streaming events.