The Real-Time vs Offline Boundary
The architecture's central dividing line: hard real-time below, latency-tolerant learning above, with telemetry crossing up and validated models crossing down.
The line that organizes everything
If the architecture has one central boundary, it is between the real-time world and the offline world. Below the line, timing is guaranteed and actions are irreversible; above it, compute is abundant and work can retry. Understanding where a function sits relative to this line explains most of its design.
Below the line
- L1 Control Plane — sub-10-microsecond determinism.
- The autonomous failsafe and deterministic interlocks.
- The soft-real-time twin shadow that informs, but does not bypass, control.
Above the line
- L0 batch retraining and Monte Carlo neutronics.
- MLOps model validation and promotion.
- Design sweeps and ecosystem scheduling.
What crosses, and how
Only two things cross the boundary, and each in a controlled way. Telemetry crosses upward as data — validated, governed, never actuating. Models cross downward only after validation, and even then their outputs are bounded by the rules engine before reaching L1. Nothing crosses the line as a live command; the offline world can shape control, but it can never be control.
Why the discipline holds
This boundary is what reconciles the two things a fusion plant needs at once: to learn continuously and to stay provably safe. Learning is placed entirely above the line so that it can never delay or corrupt the deterministic layer below. The determinism versus intelligence page argues why the plant must be built this way.
The boundary is timed by the latency gradient and drawn physically by the edge-to-cloud continuum.