Inference APIs once promised simple portability: you’d send input, get output, and own the full transcript—messages, instructions, tool calls, results. In practice, providers are locking down key pieces of state. They return encrypted “reasoning tokens,” compacted context blobs, server-stored IDs, hidden subagent messages, and opaque search results. That means your local transcript is only a partial view; the real session remains on the provider’s infrastructure. A true portable session should let you export a self-contained archive, import it into another service, and continue without hitting the original provider for decryption, cache lookups, or search results. The author proposes five tests: inspection (seeing exactly what the model saw and did), export (everything stored locally), replay (another system can pick up and carry on), audit (explaining after the fact), and deletion (removing all server-side copies). Response IDs, ciphertext blobs, and opaque summaries fail these tests. Encryption often markets itself as privacy. In reality, providers hold the keys and control where encrypted_content can be replayed. OpenAI’s store:false flag only avoids saving data on disk but doesn’t give you decryption power. Gemini’s Interactions API defaults to storing conversations for days. Anthropic encrypts full thinking in a signature field and supplies only summaries if you opt in—and warns those summaries won’t transfer between models. Hosted web search illustrates the same problem. A client-side search tool hands you URLs, passages, timestamps—enough to refetch or archive the evidence. Hosted search by OpenAI, Google, Anthropic hides the text context, offering only snippets, citations, or result counts. You see the final answer but not the actual data the model consumed. That discrepancy breaks replayability and auditability outside the provider’s ecosystem.