Determinism vs Intelligence
The architecture's core trade-off, resolved by separation: determinism where timing must be guaranteed, intelligence where compute and judgment pay off.
Two properties that resist each other
Determinism and intelligence pull in opposite directions. Determinism wants simple, fixed, verifiable logic with bounded timing. Intelligence wants large models, abundant compute, and the freedom to be occasionally wrong on the way to being usually right. A fusion plant needs both, so the architecture does not blend them — it separates them.
Where determinism wins
- Magnet quench protection and the autonomous failsafe.
- Disruption-response and plug-protection actuation.
- Vacuum interlocks and synchronized gates.
- Anything on the sub-10-microsecond command boundary.
Where intelligence wins
- Predicting the plasma state 50-100 ms ahead with the twin.
- Detecting sub-threshold quench and instability precursors.
- Planning shape or plug adjustments with MPC.
- Optimizing design and operations across large spaces.
Resolution by boundary
The trade-off is resolved by the real-time-versus-offline boundary and the safety hierarchy: deterministic functions live below, intelligent ones above, and the interface between them is a contract in which intelligence proposes and determinism disposes. Neither property is compromised, because neither is asked to do the other's job.
Why not one or the other
A purely deterministic plant could not reason about a coupled nonlinear plasma fast enough to run efficiently; a purely intelligent one could not guarantee the microsecond response the magnets demand. The architecture is the argument that you can have both — but only if you keep them apart and wire them together through contracts.
This split defines the real-time versus offline boundary and orders the safety hierarchy.