Execution
Phase 04The certified hub runs the blueprint. Fixed cells do the bounded, repeatable work they are best at; humanoids handle the reconfiguration and the flexible work between them; the factory OS controls the run and streams traceability as its default data exhaust. Nothing on the floor is improvised — everything executes a version.
Three actors, one blueprint
Execution is where Layers 03 and 04 earn their place — and where the model is honest about what is proven and what is still a bet.
The reliable core
Enclosed, bounded, repeatable tasks under controlled geometry — today's physical AI at its strongest. Proven technology, and the model leans on it deliberately. The box beats the body →
The flexible layer
Reconfiguration and the work between the cells: positioning, feeding, tending, retooling, clearing exceptions — plus simple assembly where flexibility beats fixturing. This layer is a hypothesis, tracked openly. The humanoid-race tracker →
The control layer
Configures the floor, schedules the run, enforces the quality gates, and records everything. The blueprint is what it executes; the data exhaust is what it returns.
From blueprint to running line
The factory OS translates the blueprint into this hub's reality: station assignments, tool lists, schedules, and the data contracts every machine, worker, and robot will report against. The same blueprint lands differently on different floors — the OS is what makes that translation exact instead of tribal.
The reconfiguration layer at work: humanoids move cells, fixtures, and tooling into the blueprint's layout — the physical flexibility that lets one floor change shape without a retrofit. This is the model's active hypothesis, and it is flagged as one everywhere it appears.
The first part off the line is measured against the blueprint's quality gates. Pass, and the run continues. Fail, and the deviation is logged against the blueprint version — not against someone's memory. Every deviation becomes data the next blueprint version can consume.
Fixed cells execute the bounded tasks. Between them, humanoids do the unglamorous work that keeps a line alive — loading material, swapping tools, clearing exceptions: the jobs that are pure cost while only humans can do them. The OS streams every unit's history as it happens: measurements, station, tool, operator or robot, timestamps.
| Step | Station | Executed by | Data captured | Gate |
|---|---|---|---|---|
| Seat PCBA in housing | Cell C2 | Fixed cell | Placement vision check | Pass |
| Route harness | Bench B1 | Humanoid H-03 | Task time 74 s · exception count 0 | — |
| Mount sensor module | Cell C3 | Fixed cell | Torque curve, 4 points | Pass |
| Fit seal + lid | Bench B1 | Humanoid H-03 | Torque 0.6 N·m, logged (driver T-114) | Pass |
| Functional test | Test T1 | Test station | Full record to EX-130's test spec | Pass |
Every row also carries the blueprint version, timestamps, and tool IDs. Traceability is not an audit exercise here — it is the default data exhaust of the run. The sequence matches the customer's suggestion because validation confirmed it — not because it was binding.
Nothing changes because someone moved a cart
The thesis's change-control rule applies on the floor, not just in engineering: things do not change because someone improvised a station — they change because an approved blueprint version changed. That is what keeps the floor auditable, transferable, and scalable. Physical rollback is harder than software rollback, so the system needs rollback classes — and every blueprint carries a WIP policy: complete, quarantine, transfer, freeze, or scrap, decided in advance, not in a panic.
What has to be proven
The cells are proven; the humanoid layer is not — reliability in a factory means low intervention rates, stable task success, and safe failure modes, not one-off demos. That is the model's most-watched bet. The humanoid-race page tracks its market side — price, availability, verified customers, units in the field — and independent trackers are emerging with the same discipline: Korthos's production-claims taxonomy, for instance, separates achieved output from capacity targets from marketing. The register dates the claims: P-01 (humanoids doing real, paid industrial work), P-02 (first wide adoption in repetitive, dangerous, undesirable work), P-03 (unit cost falls well below the ~USD 20,000 of 2026), and P-27 (someone publicly demonstrates a certified process transferred between robotic cells owned by two different organizations, by mid-2027). See the register →