A claim is not
the same as evidence.

Endpoint compatibility, response behavior and declared model metadata can each be examined, but they answer different questions. A plausible response is not conclusive model identity, and a successful API request does not demonstrate a confidential runtime.

Signatures can authenticate bytes and possession of a signing key. They do not by themselves prove which model ran, that the hosting operator could not inspect a prompt, or that an output is safe.

Confidential-serving goals require actual hardware and software evidence, appropriate verified measurements and a protected transport/key path. A confidential-computing label or a seller-reported file hash is not enough.

Confidential inference is a development objective, not a blanket guarantee for every model or route. This website does not demonstrate or offer a live protected-serving tier.

Training use, retention and logging policies depend on the actual route and provider. A no-training policy, zero retention and hardware-based operator exclusion are different properties and must not be treated as interchangeable.

The intended product should make relevant route information understandable and respect the user’s selected requirements. This description is a product direction, not evidence of an already accepted end-to-end implementation.