How it works
Goal → Plan → Act → Verify → Correct
Not Prompt → Action. That reads like a subtle difference; it is the difference between a demo and a business.
The sequence
Six decisions, in order.
Capability, not brand
Every body advertises what it can do: locomotion, hands, payload, precision, endurance, and discrete verbs like stairs, doors and elevators. A ranked scale cannot say that a tracked robot climbs stairs a legged one refuses — and nothing implies doors or lifts, which is the difference between working one floor of a hotel and working the building. Those must be advertised, never inferred.
Decompose against a standard
A named quality bar is a contract, so it decomposes from a template every time. Below that, the WorkGraph expands recursively — but only where no single skill can do a task whole. The depth a job needs depends on which skills exist, not on how detailed a plan looks.
Route per stage
Perception, planning, manipulation and verification are different problems and rarely have the same best model. Each is routed independently on capability match, measured field success, latency and cost — and every hop is recorded, including the ones that failed.
Dispatch by scarcity
Among the tasks a machine can do, it takes the one fewest other machines could. Otherwise the only bimanual robot takes the bathroom three bodies could clean, and the bed that only it can make waits — finishing the whole job strictly later for no reason.
Verify in the runtime first
A deterministic check is cheaper and harder to talk into a false pass than asking a model whether something happened. Offline and uncheckable returns "uncertain", which keeps the work out of the metric where it belongs.
Recover, or call a person
Working out what to do after a failure is a different reasoning problem from planning the attempt, so it gets its own stage. It must never propose the action that just failed, must diagnose a class of failure rather than an incident, and may answer "send a human" — escalating early is cheap.
Routing
Watch a different model win.
Change the body, the latency budget or the cost ceiling. The switch that matters is the last one — turn field evidence off and the cheapest model wins nearly every time, which is exactly what a router without a flywheel does.
Selected groot — capability + prior. Turn off field evidence and watch the cheapest model win instead. That is what a router without a flywheel does, and it is why the flywheel is the company rather than a feature.
Verification
Most systems have two verdicts. The third one is the point.
Click any postcondition to change what the evidence shows. Removing evidence does not produce a pass — it produces uncertain, and uncertain counts as zero.
Every postcondition was checked and held. This work counts toward Verified Autonomous Jobs.
Click any condition to change what the evidence shows. Most systems have two verdicts; the third one is why the metric is trustworthy.
Position
Every layer but ours is being driven to commodity.
That is the argument, not a caveat. We route across whatever is best today and keep the record of what actually worked — which is the only asset in this stack that compounds rather than depreciating.