The original promise of inference APIs was simple: send input, receive output, keep the conversation. That abstraction is now fragile because prompt caches, tokenization differences, and non‑reproducible sampling create provider‑bound state.

Even a transcript that contains instructions, messages, and tool calls is now often an opaque blob. Providers return encrypted content, hidden searches, and server‑side context that only they can decrypt. The result is a partial view that cannot be fully replayed elsewhere.

Portability should mean a self‑contained export that a new model can continue from without needing the old provider’s keys or IDs. Five tests—inspection, export, replay, audit, deletion—highlight how current APIs fall short. IDs, encrypted blobs, and citation lists are not user‑controlled, so the session remains tied to the provider.

Providers market this as a privacy feature, but the encryption actually hides data from the user. Stored conversations, provider‑sealed state, and hidden searches keep the logic inside one ecosystem, making true portability impossible.