SENKU.IM
Running experiment

Model fusion · one household

Two minds. One checked result.

The model-fusion loopClaude dispatches a task to a Codex worker, cross-checks its result, and verifies the artifact before shipping. INPUTTask EXECUTESolCodex worker ORCHESTRATEOpusClaude OUTPUTShip CROSS-CHECKRead the real artifact
Claude dispatches. Codex executes. Claude judges. The system verifies what was actually made.
Current stage

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.

Fusion Studio3 workers active
Illustrative system view. The roles, workload, budget, and artifacts are generic; no private job data appears here.
  1. Queue
    work waits
  2. Claim
    one worker owns it
  3. Isolate
    clean workspace
  4. Converse
    brief and feedback
  5. Commit
    every turn persists
  6. Outbox
    artifact returns
  7. 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.