Which provider served this model request?

Pin a provider, record the route and read the evidence behind a model label before comparing two API responses.

Suppose a support assistant gives a different answer on Tuesday than it did on Monday. Before spending an afternoon rewriting the prompt, check whether the two requests reached the same deployment. The model ID, provider and generation settings are a good place to start.

Make one request reproducible

Choose an explicit model ID from the model catalog. On its endpoint list, pick one available provider. Restrict the request to that provider with provider.only and turn off fallback for the comparison. Save the input, model ID, settings and timestamp with your test case. The routing reference describes those controls.

Use a short example whose answer you can check. For a support assistant, two lines of fictional shipping policy and a question about the return window are enough. If the answer varies across repeated requests on the pinned route, you have learned something before introducing a second provider.

Keep a record of the response

When you need a signed record, request an inference receipt and verify it using the receipt instructions. It ties the request and response to the route recorded by the gateway. A claim about the exact model weights running upstream needs evidence from the serving workload itself.

When comparing two providers, keep the receipt beside each provider’s published precision information. Record the review date and any unknown fields. That gives the next engineer a concrete starting point for investigating a change in behavior.