The harness
The model isn’t the product. The harness is.
Two people with identical access to the same frontier model get wildly different output — and the gap has nothing to do with prompt cleverness. This is the operating layer I run Claude on: routing, delegation contracts, review gates, and a job contract for automation nobody is watching. The whole thing is published, MIT.
Routing · by task shape, not habit
$ route --by task-shape
Intelligence compounds at exactly two points: deciding what to do, and checking whether it was done. Everything between is execution, and execution is cheap.
The loop · every delegated build
Plan
A frontier session writes a delegation contract — numbered steps, exact paths, named error behaviors.
Execute
A mid-tier model runs the plan step by step. No inferring intent — the plan is literal by design.
Verify
A review gate in fresh context runs the done-condition itself. “Tests pass” is a claim. The exit code is evidence.
Encode
Every correction lands in config the same session. A preference stated in chat evaporates; a contract template compounds.
The job contract · rules for automation nobody is watching
Guaranteed terminal status
Every job ends by reporting success or failure. Nothing is allowed to die silently.
Cost tagged at the source
Every run records what it cost and which account paid for it.
Silence gets detected
Heartbeats and outcome checks catch jobs that stop producing results, not just jobs that crash.
Test traffic is separable
Test runs are labeled so they never mix into production metrics.
The whole harness is public.
Routing policy, delegation templates, the review-gate agent, and the nine rules for unattended automation. MIT licensed.