Stage
One step of the build — form, weld, enamel, foam, assemble, test, pack — owned by one agent.
Full reference opens with design-partner access
Documentation is organised the way a deployment actually happens: connect, baseline, advise, control, then expand. Each section maps to a stage of the trust model rather than to an org chart.
The first hour, the first day and the first fortnight.
Provision K3s on site, register the plant, verify GPU availability and confirm fail-safe stop paths.
02Define the site, line, stages and product families, then attach the units the line actually builds.
03Run shadow mode for a minimum 14-day window and publish the measured baseline for your chosen metric.
Protocol, direction and the authority each connector can be granted.
| Family | Protocols | Direction | Authority |
|---|---|---|---|
| Forming / stamping press | OPC UA, PLC tags | Read + write | Parameter adjustment within a bounded envelope |
| Paint / enamel line | OPC UA, REST | Read + write | Coverage, film build and cure setpoints |
| Foaming machine | OPC UA, PLC tags | Read + write | Shot weight, mix ratio, cure time |
| Assembly stations | MQTT, REST, tool APIs | Read + write | Gating, torque targets, station instructions |
| Leak / functional testers | MQTT, REST, file drop | Read + write | Test sequencing and disposition |
| Appliance MES / quality | REST, file drop | Read + write | Genealogy, non-conformance, disposition |
Everything in the API hangs off the unit. A serial number is the join key between perception, process, test and warranty.
Factory — a site and its edge runtimeLine — stages, product families, autonomyUnit — one serialised applianceVerdict — decision, risk, citationsAuditRecord — immutable, exportable# Register a line and stream a unit through the loop
from duromex import Factory
factory = Factory(site="plant-ankara-2", edge="local")
line = factory.line(
id="refrigeration-cabinet-01",
stages=["form", "weld", "enamel", "foam",
"assemble", "test", "pack"],
autonomy="shadow", # shadow | advisory | control
)
unit = line.unit(serial="RF-2026-0004417")
unit.attach("vision", stream="opcua://press-04/cam")
unit.attach("pressure_decay", stream="mqtt://tester-11/decay")
verdict = unit.evaluate()
print(verdict.first_pass, verdict.risk, verdict.citations)
# True 0.03 ["recipe:foam-v12", "spec:IEC-60335-2-24"]Read these once and the rest of the documentation is mostly detail.
One step of the build — form, weld, enamel, foam, assemble, test, pack — owned by one agent.
The fused sensor evidence attached to a specific unit at a specific stage.
A decision with a confidence, a probable cause and a citation chain back to a source.
What an agent may write, on which line, within which bounds — granted, scoped and revocable.
Shadow, advisory or control, set per line and per workflow.
The simulation check a change must pass before it can reach production.
The immutable log entry for every action: evidence, reasoning, write and approver.
The measured human and tester performance the agent must beat, captured in shadow mode.
An engineer's override — the highest-value training signal the system receives.
Review exceptions the agents held for human decision, and confirm the overnight drift and evaluation status.
Review the metric trend against baseline and inspect the corrections your engineers made most often — they point at the next model improvement.
Rehearse the model change in the twin, confirm the predicted yield and takt, then release under canary.
ROI review against the written success metric, and a decision on expanding autonomy, modules or lines.
Each stage of the build has its own connector reference, model card, control envelope and evidence schema — so a question about foam is answered in the foam section, not in a generic API page.
Nine steps, in this order, every time.
Choose one line, one workflow and one written success metric with a named owner.
Install the runtime on site and verify GPU, network and fail-safe stop paths.
Register the connectors for that line's press, enamel line, foamer, stations, testers and MES.
Run shadow mode for at least fourteen days and publish the baseline and accuracy gap.
Enable advisory mode with human approval on every action; corrections train the models.
Wire the twin and CI evaluation gates into your change process before any control is granted.
Grant scoped, revocable write authority for low-risk workflows and escalate exceptions.
Run the quarterly ROI review against the written metric.
Add the adjacent module, then the next line, then the next site.
Stated plainly — Duromex is pre-launch.
sections outlined publicly
access tier today: design partners
of connector families documented at pilot
public API keys issued yet
Through a design-partner engagement or a paid pilot. The full reference includes control envelopes and safety-relevant detail that we do not publish openly during pre-launch.
Yes, provided with access. The agent API and the audit API are both specified, and the audit schema is versioned additively.
Each stage model ships with a card describing training data provenance, evaluation results, known failure modes and the confidence thresholds used for escalation.
Design partners and pilot customers get the full reference, the SDK and the edge runtime.
Duromex is pre-launch. Figures shown are design-partner targets and modelled economics, not audited results. Ask us for the methodology.