> For the complete documentation index, see [llms.txt](https://docs.trezalabs.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.trezalabs.com/ai-gateway/attestation.md).

# Attestation

On the Enterprise plan, redaction runs inside a **Treza Enclave** (AWS Nitro) instead of the managed software pipeline. Hardware attestation lets you cryptographically verify *where* your prompts were redacted — before any of them left your trust boundary for the LLM provider.

{% hint style="info" %}
TEE-attested redaction is an **Enterprise** entitlement. Starter and Pro accounts always use managed software redaction (`x-treza-mode: standard`) and never route through a Treza Enclave. See [Plans & Pricing](/ai-gateway/plans-and-pricing.md).
{% endhint %}

### Redaction modes

| Mode       | Plans        | Where redaction runs                         | Attested |
| ---------- | ------------ | -------------------------------------------- | -------- |
| `standard` | Starter, Pro | Managed software redaction                   | No       |
| `tee`      | Enterprise   | Treza Enclave (AWS Nitro), hardware-isolated | Yes      |

The mode is reported on every request:

* `POST /api/redact/chat/completions` — the `x-treza-mode` response header
* `POST /api/redact/run` — the `mode` field, plus an `attestation` object (`enclaveId`, `region`, `attested`) on Enterprise
* `GET /api/redact/log` — each audit entry's `attestationRef` records which environment performed the redaction (`shared-software-redaction` for standard mode, the enclave id for TEE mode)

{% hint style="warning" %}
`x-treza-mode: tee` is only set when the redaction actually ran inside an attested enclave. If the enclave cannot be reached, the proxy fails closed (`502 redaction_failed`) rather than silently downgrading — your unredacted messages are never forwarded upstream.
{% endhint %}

### What the attestation proves

A Treza Enclave produces a signed attestation document with Platform Configuration Register (PCR) measurements. For the redaction service this proves:

* **Code identity** — the exact redaction enclave image (`pcr0`), kernel and bootstrap (`pcr1`), and application (`pcr2`) that processed your text
* **Isolation** — redaction ran inside a hardware-isolated Nitro Enclave with no operator or host access to the plaintext
* **Engine versions** — the `modelVersion` and `recognizerVersion` of the redaction engine, matching what is stamped on your audit entries

It does **not** attest the upstream LLM provider — only the Treza redaction step that runs before your request is forwarded.

For background on PCR measurements and attestation documents, see [Attestation & Verification](/getting-started/concepts.md#attestation-and-verification).

### Get the attestation summary

`GET /api/redact/attest` — requires a Treza API key with the `redact:run` scope.

```bash
curl https://trezalabs.com/api/redact/attest \
  -H "Authorization: Bearer $TREZA_API_KEY"
```

```json
{
  "enclaveId": "eg-0a1b2c3d4e5f6a7b8",
  "region": "us-west-2",
  "pcr0": "e3b0c44298fc1c149afbf4c8996fb924...",
  "pcr1": "9f86d081884c7d659a2feaa0c55ad015...",
  "pcr2": "2c26b46b68ffc68ff99b453c1d304134...",
  "modelVersion": "1.4.2",
  "recognizerVersion": "2.1.0",
  "attested": true
}
```

| Field                    | Meaning                                                                                 |
| ------------------------ | --------------------------------------------------------------------------------------- |
| `enclaveId`              | The Treza Enclave performing your redaction (matches `attestationRef` in the audit log) |
| `region`                 | AWS region the enclave runs in                                                          |
| `pcr0` / `pcr1` / `pcr2` | PCR measurements of the enclave image, kernel, and application                          |
| `modelVersion`           | Redaction model version running in the enclave                                          |
| `recognizerVersion`      | Entity recognizer version                                                               |
| `attested`               | `true` when the summary comes from a verified enclave attestation                       |

On non-Enterprise plans the endpoint returns `403` with `code: "PLAN_FEATURE"`:

```json
{
  "error": "Hardware attestation requires an Enterprise plan.",
  "code": "PLAN_FEATURE"
}
```

### Tying attestation to individual requests

Every redaction request — proxied or standalone — writes an audit entry whose `attestationRef` is the id of the environment that redacted it. To build an auditable chain:

1. Capture the `x-treza-request-id` response header on each proxied request.
2. Look the request up in `GET /api/redact/log` (or the [CSV export](/ai-gateway/usage-and-billing.md#audit-log-export-pro)); its `attestationRef`, `modelVersion`, and `recognizerVersion` identify the redaction environment.
3. Compare `attestationRef` against the `enclaveId` from `GET /api/redact/attest` and verify the PCR measurements against your approved values.

### Related

* [Quickstart](/ai-gateway/quickstart.md) — send your first redacted request
* [Usage & Billing](/ai-gateway/usage-and-billing.md) — audit log export
* [Attestation & Verification concepts](/getting-started/concepts.md#attestation-and-verification)