Five checks from the local experiment
- 1. Bind locally: keep the first endpoint on 127.0.0.1 while behavior and credentials are still under examination.
- 2. Verify authentication rejection: the unauthenticated model-list request returned 401 in this setup.
- 3. Verify authenticated discovery: the authenticated model-list request returned 200, with one imported free account active.
- 4. Send one minimal completion: the auto request returned 200 in 9.92 seconds and used 42 tokens.
- 5. Record limits and failure boundaries: this service used local SQLite, while proxy runtime and clearance were disabled.
Compatibility is not an operational guarantee
- The five checks establish only the observed authentication, discovery, storage, and completion behavior of this local run.
- An OpenAI-compatible request and response shape is not evidence that an endpoint is the official OpenAI API or that it is reliable enough for production.
Questions this lab note should answer
Why start with a localhost binding?
It keeps the unproven service off public interfaces while authentication and response behavior are tested.
Why test an unauthenticated request?
A protected endpoint should reject the same request when its credential is absent; this test returned HTTP 401.
What is the final pass condition?
A minimal completion should return a usable response, but that pass remains a smoke-test result rather than a production guarantee.
The generator and the evidence boundary are disclosed.
The first draft was generated through the locally hosted chatgpt2api experiment using only the measured test facts listed here. FreeToken removed the generator's wrapper text and manually checked the retained statements before publication.