Start with one useful process
One manufacturing workflow, explicit acceptance criteria, known authority, and a measurable operating result.
A process is defined as a fixed, versioned sequence of steps. Perception and geometry describe the part, planning and control carry out each step, and a set of checks has to pass before anything moves. This page walks through that sequence and the boundaries around it.
One manufacturing workflow, explicit acceptance criteria, known authority, and a measurable operating result.
Each node can use a learned model, CAD operation, planner, rule, controller, or human decision.
Every transition has an acceptance contract, recovery route, and evidence trail.
THE VERSIONED EXECUTION GRAPH
The orchestrator reads state and confidence, selects the next approved skill, evaluates authority, and either executes, recovers, or stops. Learned models never become the safety or quality system.
{
"objective": "finish_to_tolerance",
"part_id": "NI-24-00841",
"recipe": "MED-NITI-014@6",
"surface_delta_um": 3.8,
"tool_state": "ready"
}
{
"decision": "EXECUTE_RECIPE",
"confidence": 0.997,
"allowed": true,
"gates": "8/8",
"action": "finish_pass_03"
}
ARCHITECTURAL BOUNDARIES
FROM PROCESS TO EXECUTION GRAPH
Get one physical process working under a clear operating contract, then improve and reuse only what production evidence supports.
Capture the manual baseline, representative cases or parts, limits, quality criteria, and failure modes.
Build the graph: state estimates, skill nodes, transitions, authority, recovery paths, and acceptance tests.
Run the graph against representative real parts. Collect synchronized state, decisions, actions, and outcomes.
Test each node and the full graph, freeze the approved version, train operators, and establish change control.
Improve weak nodes from evidence and reuse validated skills across adjacent parts, processes, and sites.
SYSTEM READY
Enter a demonstration of state estimation, orchestration, modular skills, authority gates, verification, and recovery.