Docs for Fabric Verify

Send a request to the open model we host (qwen2.5:3b). An observer in an Intel TDX trust domain of its own makes the call and returns a signed receipt. That machine runs only the model and the observer, on an OS image with no SSH, so nobody, including us, can log in to it. The receipt shows that a genuine trust domain matching our published measurements held the key that signed it. It also shows that the trust domain measured the model files before that key existed, and that this exact request and answer went to that model. Each run costs $0.05 of Fabric Verify credit, charged only on success.

GET /v1/attested-execution/attestation shows the current boot's key attestation and the measured model. pins/README.txt shows how to build policy.json from our published pins. A receipt does not show which weights computed each token, anything about GPUs, or that the answer is correct. We can still deploy a different configuration to that machine, but it would change the published measurements, so receipts from it would fail your pinned policy.

Each run is also recorded in your log and sealed a few minutes later, at no extra cost. Fetch the run again with GET /v1/attested-execution/runs/{run_id}: once the x-fabric-log-status header says sealed, the receipt carries its log evidence, and the verifier also reports log inclusion. That record is signed by a key we hold (labelled hosted). pins/README.txt has the steps, including recording it with your own key instead.

Set a policy for your account, and the log applies it when your agents record actions. It can limit which keys may record and which kinds of action, require numbered records, and require an approval before chosen kinds of action. An approval is a record from an approver key you name. The log lets it cover one gated action, once, within the time you set. You can also halt one agent key, or your whole account, and resume it later. A refused record is not charged and not logged; the refusal is kept for your governance export. Everything here is free except the export.

A refused record gets HTTP 403 with governance_refused, the rule and a reason code, for example approval_missing, approval_used or agent_halted.

Also free:

• GET /v1/governance/policy and GET /v1/governance/state: the current policy and the halts in force.
• GET /v1/governance/events?since=<RFC 3339 time>: a feed for your SIEM (one JSON object per line) of policy changes, halts, resumes, approvals and refusals. It is not signed; the export is the signed record.
• GET /v1/log/tenant-head: the log's signed count of your account's records.

The export ($1, the audit-pack price) is signed by the log. It holds your policy versions, your halts, your approvals with the record that used each one, and your refusals. If your account has up to 1,000 records, it carries every proof, so the verifier can show that none is missing. It does not show that the service recorded every refusal, actions your agents never sent to the log, or that a policy or an approved action was right. The verifier also checks an account's or a key's full set of records, and bundles of linked agent events: see CLI.txt.

A proof shows a record was in the log, unchanged, by a certain time. It does not show the record's content is true. Records hold a fingerprint (hash) of the data, never the data.

Check it yourself

Download the verifier and our public keys at augmentev.com/verify. The security page explains the limits.