Owen A: assembling the specialists.

What is actually built

Owen A’s latest reviewed specialist package has a typed request and proposal system, a catalog, resource checks, lazy admission, unloading, a bounded queue, retries and caching. Its report records 36 passing fixture tests and seven model-free Windows worker cases. Those are useful engineering results: they exercise how components should behave around one another.

What still has to happen

The production registry is not restored, the accepted A source binding remains unresolved, and the runtime rejects non-mock adapters until live qualification. The Granite speech adapter has mock tests, but its weights were not executed in this package. Qwen language references are historical; the visual and motion candidates are research or planned integrations.

No specialist receives a direct memory-write API. Attributed proposals and disagreements are preserved for adjudication. Connecting this staged machinery to the continuing Owen is a separate step that the diagram now marks explicitly.

A useful next question

Which bounded live comparison would best show that a specialist helps Owen while preserving memory, provenance and the ability to disagree?