# How to verify which provider answered your LLM request (/blog/verify-llm-provider)

Turn on a signed receipt for an API key, check it yourself, and learn what the receipt proves and what it does not prove.

When you call a model through another company's API, you cannot see the connection behind it. A signed receipt lets you check one part of that connection: the provider host that received your request, and the exact bytes that went in and came out.

OhMyGPT can attach such a receipt to a chat request. [VeriRelay](https://verirelay.com) signs the receipt with a key that exists only inside a confidential virtual machine, and it publishes the attestation for that machine.

## Turn on receipts for a key [#turn-on-receipts-for-a-key]

1. Open the [API keys](https://www.ohmygpt.com/keys) page.
2. Create a key, or edit an existing key.
3. Switch on **TEE confidential computing proof**.

Your code stays the same. Every request that you make with this key asks for a receipt.

## Read the response headers [#read-the-response-headers]

A response with a receipt carries three headers:

```http
X-VeriRelay-Request-Id: 36a420474e713c4ecbe0a6d5dfd79ed0
X-VeriRelay-Store-Id: cFMrwmc4IeTQT14_R9y6gg
X-VeriRelay-Store-Key: <secret>
```

Treat the Store-Key as a secret. Anyone who holds it can read that request and its response until VeriRelay deletes them.

A stream can carry the request ID and the store ID as comment lines instead of headers. [TEE Proofs](/docs/tee-proof#what-a-proven-response-carries) describes that case. If a response has neither the headers nor the comment lines, it has no receipt. A request that cannot get a receipt is still answered, so check for these values when your process depends on them.

## Check a receipt in the browser [#check-a-receipt-in-the-browser]

1. Open the [Usage](https://www.ohmygpt.com/usage) page.
2. Open a request that is marked **Verifiable**.
3. Click **Verify on VeriRelay**.

The page checks the signature and the attestation. It then decrypts the stored bodies in your browser and compares them with the hashes in the receipt.

## Check a receipt in code [#check-a-receipt-in-code]

Fetch the receipt with the request ID. The call needs no key:

```bash
curl https://relay.verirelay.com/v1/avr/receipts/36a420474e713c4ecbe0a6d5dfd79ed0
```

Three fields answer the main question:

| Field | Meaning |
|---|---|
| `upstream_authority` | The host that answered, such as `api.deepseek.com` |
| `request_body_sha256` | The hash of the request body that the host received |
| `response_body_sha256` | The hash of the response body that the host sent |

VeriRelay keeps a receipt for 1 hour and the stored bodies for 24 hours. To keep a proof longer, save the evidence bundle within the hour:

```bash
curl -o evidence.json https://relay.verirelay.com/v1/avr/evidence/36a420474e713c4ecbe0a6d5dfd79ed0
```

## Know what the receipt does not prove [#know-what-the-receipt-does-not-prove]

The receipt proves where the request went and what bytes were exchanged. It does not prove which model the provider ran behind its own API, and it says nothing about the quality of the answer.

It also covers the exchange with the provider, not every byte that you receive. OhMyGPT changes a few fields, such as the model name, and [TEE Proofs](/docs/tee-proof#what-the-receipt-covers) lists each change. Embeddings, images, audio, and some chat models have no receipts.
