L3 Contrasted: Breeder vs Burner

The same L3 framework instantiates differently for a spherical tokamak and a tandem mirror; here is what changes and what stays the same.

One framework, two physics

Both machines run the same L3 architecture, GNN diagnostics, PINN physics, anomaly ensembles, MPC, and the KRONOS-CTRL twin with a 50-100 ms shadow, but the physics they model and control is different, so the instantiation differs in specific, well-defined ways.

What stays the same

The shared parts are substantial: the L2 feature conventions and lineage, the sensor-topology GNN and its imputation, the surrogate-acceleration and UQ machinery, the state-estimation and shadow-synchronization loop, the anomaly-ensemble framework, the MPC formulation with a certified envelope and terminal set, and the V&V, calibration, drift-monitoring, and confidence-scoring discipline. Only the physics modules and the control targets are swapped.

This is deliberate. Building one rigorous L3 framework and instantiating it twice means the breeder and burner programs share validation machinery, transfer learning, and operational tooling, and lessons from the breeder (which reaches FOAK first, ~2030) inform the later burner without rebuilding the stack.

Both remain design/simulation studies pre-FOAK, with no hardware net-gain claimed, and both inherit the same strict layering: L3 governs performance within a certified envelope while the L1 hardware failsafe guarantees safety independently of any model.