Model fusion · one household
Two minds. One checked result.
A real running household system. The studio below is an illustrative system view, built from generic roles and aggregate activity. It is not a live feed or a product demonstration.
One system, two altitudes
At the model level, Opus orchestrates Sol workers and cross-checks their output. At the engine level, jobs move through a queue into isolated workspaces, then through a conversation where each turn becomes a durable commit. The result returns through an outbox for judgment.
The second view is the machinery that makes the first repeatable.
The orchestration floor
Several Codex workers can execute in parallel while one orchestrator keeps the brief, reviews each checkpoint, and decides what ships.
- Queue
work waits - Claim
one worker owns it - Isolate
clean workspace - Converse
brief and feedback - Commit
every turn persists - Outbox
artifact returns - Judge
verify and ship
What the engine protects
Isolation keeps parallel jobs from writing over one another. A commit after every turn makes the conversation recoverable. The outbox gives the orchestrator a concrete artifact to inspect. Cross-checking asks whether the saved result works, beyond whether a worker reported success.
A failure it caught
On 12 August a build step reported success while its saved output was cut off at a fixed buffer ceiling. Checking the artifact found the exact-length truncation. The capture path now writes to disk without that ceiling, and a regression test rejects the old failure.
A watcher that drafts operator corrections exists, but has not yet fired in live work.
Flagship demonstration
A Russian book was translated into English. Each model produced a merged version, then one judged both anonymously against a rubric agreed in advance. The winning version scored highest on voice.
Read the translation demonstration →Related writing
Loading the blog’s published posts…
Honest limits
Three limits to keep in view
This is one household system, so it does not establish a general result. It uses more compute because several workers and a judging pass can participate. Models can still share the same blind spot or follow a weak rubric.