The bounded MCP prototype established an important feasibility boundary: a private appliance can preserve a versioned memory snapshot and retrieve a bounded, source-linked evidence packet with provenance and later correction or limitation boundaries.
That packet was delivered to a fresh Codex task, which remained responsible for the final answer. Earlier local-model qualification and fallback experiments are not presented as a customer-facing answer service or an external-AI handoff.
A compelling future is not evidence that the future has already shipped.
The status ledger
The public website now separates demonstrated capability from active development and product roadmap. That distinction will move with the build receipts, not with enthusiastic conversation.
The current Bob and model-control-room work is a source candidate and synthetic shadow. It does not yet mean customer-ready provider routing, tools, autonomous work, or a shipping appliance.
Why the boundary matters
A memory system will eventually hold material that matters deeply to its owner. Trust cannot be built from words like private, sovereign, or intelligent alone. It has to be earned through permissions, receipts, restore tests, safe failure, and visible limits.
The goal is ambitious. The reporting can still be exact.