The knowledge worker's goal — "when I search for it, I want today's version" — is the entire justification for the interactive freshness lane and for atomic version visibility.
The tenant admin's goal is the reason revocation is on the read path and not on the pipeline's freshness path: five seconds, not ten minutes.
The platform engineer's goal is the reason the critical design decision exists at all. Everything about contracts, aliases and dual-write is downstream of it.
Machines are actors too
The change feed is in the cast because its goal — "be told plainly when I am being throttled" — is a requirement: typed, retryable rejection rather than a silent drop.
The reconciliation sweep exists because the feed will miss documents, and the only alternative to a sweep is a user finding the gap.
The quality harness is an actor because somebody has to be able to say retrieval got worse, and no human watches 4.8 billion chunks.
Deliberately not here
The answer-layer prompt engineer: a real role, but they consume retrieval and do not change this architecture.
The data scientist choosing the embedding model: their output is a model digest, which enters as an input to a contract.
