9 hours ago
- Inference APIs are moving from simple, portable transcripts to provider-bound, non-portable state, using encrypted reasoning, hidden search results, opaque compaction, and encrypted agent messages, locking users into a single ecosystem.
- Session ownership is tested by five criteria: inspection, export, replay, audit, and deletion, which most current APIs fail because they rely on server-side IDs, encrypted blobs, and unresolved references.
- Features like 'encrypted content' are misleading; they hide data from the user, not the provider, and are better termed 'provider-sealed state' for privacy theater.
- Server-side storage of conversations (e.g., OpenAI's Responses API, Gemini Interactions API) turns transcripts into pointers, making them non-portable and dependent on the original provider.
- Missing raw reasoning tokens and encrypted reasoning blocks prevent portability; users cannot inspect or transfer reasoning to another model, even within the same provider.
- Hosted search lacks full-fidelity logs, providing only citations and not the complete context (queries, rankings, passages) used, breaking transcript completeness.
- Opaque compaction (e.g., OpenAI's encrypted compaction) hides meaning from the user and other providers, unlike Anthropic's readable compaction which allows inspection and transfer.
- Subagent communication in multi-agent systems is encrypted, hiding instructions and messages from the application and making audit and portability impossible.
- The default 'store: true' and lack of explicit storage options create lock-in, while portable APIs should mandate explicit storage, readable handoff representations, and full-fidelity logs.
- Distillation should be supported, not treated as an attack, to enable local, offline, and user-controlled models, reducing lock-in and increasing competition.