Getting Started

llm-id checks whether an API endpoint really serves the model it claims to serve. Relays and resellers sometimes swap in a cheaper model, serve a degraded copy, send part of the traffic elsewhere, or sell a copy with its refusals removed. llm-id asks the endpoint short questions many times and compares the answers with answers collected from the model's official provider.

What a Test Can Tell You

Every test ends with one verdict and a confidence rating:

• Genuine: it's the claimed model.
• Degraded: the claimed model, but measurably weaker than the official deployment.
• Abliterated: the claimed model with its refusals removed.
• Substituted: a different, named model answered.
• Unknown Model: no model in the baseline fits the answers.
• Mixed Routing: requests go to several models.
• Inconclusive: not enough evidence either way.

Reading a Report explains each verdict and what to do about it.

Your First Test

1. Select Sign In at the top of any page, then Create an Account, and enter your email address and a password of at least 8 characters. Next time, sign in with them.
2. Select + Test Endpoint in the header.
3. In Endpoint Base URL, enter the address the endpoint gives you, such as https://relay.example/v1. The field shows an example for the chosen API.
4. In Model ID at the Endpoint, enter the model name the endpoint expects. List asks the endpoint for its model list, and you can pick from it.
5. In Claims to Be, pick the model the endpoint says it serves.
6. Paste your key into API Key.
7. Leave API, Tier and the options as they are for a first test.
8. Select Start Test. The page follows the test as answers arrive. Most tests take a few minutes; reasoning models take longer.
9. Read the verdict at the top of the report, then the sections below it.

Testing an Endpoint explains every field and option.

What a Test Costs

Every request goes to the endpoint on your key, so the endpoint bills it as usual. The side panel shows the count before you start.

Identity and refusal probes are short questions. Competence items are the expensive part: reasoning models spend hundreds to thousands of tokens on each. That's why the competence and refusal checks are off unless you switch them on.

Your Account, Key and Results

• Every test is saved to your account. It's stored on our side with its full report, so it stays available from any device you sign in on. It appears under Checks, where you can reopen the report, filter by verdict and export it. Only your account can see it.
• Your key is used for one test only. It's held in memory while the test runs, and it's never stored or written into a report. To rerun a test, enter it again.
• Your password is stored only as a salted hash. A sign-in lasts 30 days on a device; Sign Out in the header ends it sooner. After 10 wrong passwords in a row, sign-in pauses for up to 15 minutes.

Before You Act on a Verdict

A verdict other than Genuine is evidence, not proof of intent. Confirm it with a separate test at the Thorough tier, and read Limits first.