Model compatibility lab: turn a familiar API shape into a tested capability matrix

Begin with the documented contract

UncensuredAI documents Chat Completions, Bearer API-key authentication and the base URL https://uncensuredai.com/v1. A complete request targets /chat/completions beneath that base. Use the exact catalog model ID and avoid adding /v1 twice.

This establishes a documented integration path. It does not establish complete OpenAI Responses API compatibility. Tool calling, JSON output constraints and image inputs should be checked for the particular model, parameters and client version you intend to use. Leave unverified capabilities marked unknown rather than treating an accepted connection as proof of everything.

Build a small matrix with explicit acceptance criteria

Use one row per exact model ID and one column per required feature. Start with a basic text response, streaming completion and error reporting. Add tool calls, structured JSON or image understanding only if your product needs them and the model’s current documentation supports testing them.

Define success at the application level. A JSON test should check parsing and your required fields. A tool test should check the returned name and arguments against a harmless fixture tool. An image test should ask about a controlled image whose expected answer you know. HTTP success alone does not establish feature correctness.

Separate parser checks from model checks

Run local fixtures through the downloadable Node and Python clients first. These can expose broken chunk handling, malformed-response handling and incorrect association of batch results. They cannot prove that a live model supports a feature or follows its instructions reliably.

For a live check, change one variable at a time and preserve the request settings with the result. Record the date, model ID, client runtime, feature tested and observed outcome. Keep private prompts and credentials out of a published matrix. Retest the affected cells when your integration changes.

Keep edition and operational assumptions visible

Original models retain producer safeguards; Uncensored and Abliterated editions are separate catalog choices. Do not copy a capability result between editions without testing it. Model context limits, output limits and shared-capacity behavior also need their own checks.

UncensuredAI’s zero-retention statement concerns prompt and response copies on its servers. Account data is separate; this published statement is not an independent audit. Your test logs and connected application can store data independently. If you resell the resulting integration, standard keys can be used at public rates without prior agreement; dedicated rates, master keys and service commitments require an agreed collaboration.

Adopt only the capabilities your evidence supports

Real API checks can consume paid credit at the chosen model’s rates. Use single attempts and do not automatically repeat ambiguous timeouts or partially completed streams. Download the clients, pass the local fixtures and test one required capability at a time. Use the catalog and API reference to keep the matrix tied to current documentation.