Sigstack verifier

Send the bot !verify <random text> in Signal. It answers with two messages: a summary with your challenge, its report data and its compose hash, then its TDX quote on its own.

1. Signature. Long-press the second message, copy the quote, and paste it into proof.t16z.com. It checks that Intel signed it.
2. Freshness. The bot quotes Sigstack: <your text> as its report data, so only code in this TEE could have produced this quote for your text. Check it below.
3. Code. The compose hash must be one the owner approved (listed under Approved builds), and it should match the registered bot above.
4. Models. The reply also reports the NEAR AI models the bot uses. The page can't see those checks; how the bot checks NEAR AI explains them.

Your text (what you typed after !verify) Report data from the reply (optional) Compose hash from the reply

Everything here runs in your browser; nothing you type is sent anywhere.

Check the report data from a terminal

The report data is the bytes of Sigstack: <your text> as hex, or their SHA-256 when they're over 64 bytes. In the quote itself, that value is padded with zeros to 64 bytes.

printf '%s' 'Sigstack: <your text>' | xxd -p (or | sha256sum for long text)

How the bot checks NEAR AI

Translation and transcription run on NEAR AI Cloud models in GPU TEEs. Inside its own TEE, the bot checks each model's attestation: an Intel TDX quote, and NVIDIA's verdict on the GPUs. It then checks that every translation is signed by the key of an attested instance. The !verify reply shows each model's signing key, compose hash, Intel TCB status and GPU count, and how many translations verified.

Whisper is attested too, but NEAR AI doesn't sign individual transcripts yet. Details: docs/model-verification.md and NEAR AI's verification docs.

Checked by the contract on NEAR, when the bot registered

• The TDX quote is signed by Intel, using collateral that is in date, and the platform's TCB is up to date.
• The quote's report data is the agent's own account ID, so a quote can't be reused by another account.
• The firmware, OS and compose-hash measurements match a build the owner approved.
• The CPU's PPID is on the owner's approved list.

Checked by this page, in your browser

• Each approved compose hash is the sha256 of a published app_compose file.
• That file pins every container image by digest.
• For this repo's images, what GitHub's signed build provenance says: the commit and workflow run that built each digest.
• The contract's on-chain code is the wasm built by this repo's CI.

Checked by this repo's CI, on every change to the build archive

• Each attested compose file is docker/phala-demo.yaml byte for byte, at the commit that pinned it.
• The images were built from that commit or an earlier one.
• On ghcr.io, the tag sha-<commit> of each of this repo's images is the pinned digest.
• Where an image has GitHub provenance, its signature verifies and it names the stated commit.

Not verified here

• Provenance signatures. This page shows what the attestation claims; run the gh attestation verify command shown next to an image to check its signature. Builds archived before provenance signing have only the tag check. Rebuilding an image yourself is the strongest check, but the builds aren't bit-for-bit reproducible yet.
• That the bot you message in Signal is this agent. Send it !verify <random text> and check its reply.
• The AI side. Messages are transcribed and translated by NEAR AI Cloud models in GPU TEEs. Builds from 5f0f9eb9 on check those models' attestation and translation signatures inside the bot, and !verify reports the result; this page can't see that.
• The firmware and OS measurements (MRTD, RTMR0–2) are Phala's dstack image. They're approved as published, not rebuilt here.