Regime Drift Detection on Plasma State

Regime monitors detect when the machine crosses into a plasma operating regime outside a model's validated envelope and gate that model accordingly.

Staying inside the validated envelope

Every deployed model has an explicit validity envelope: the region of operating space over which it was trained and validated. Regime drift detection watches the live plasma state vector and asks a single question each cycle: is the machine still inside the envelope of the model currently relying on this state? Outside it, the model's predictions are extrapolations and must not carry authority.

For the breeder this envelope is expressed in confinement and stability coordinates (normalized beta, q95, Greenwald fraction, triangularity around the canonical negative delta -0.30). For the burner it is expressed in mirror-confinement coordinates (plug field toward the 26.49 T design point, mirror ratio, ambipolar potential). Crossing a boundary is not an error; it is expected as campaigns explore. The response is to gate, not to force the model.

Envelope representation

• Convex hull / one-class SVM around training-set operating points
• Mahalanobis distance from the training centroid in state space
• Per-axis hard limits derived from physics constraints
• Density estimate flagging low-support regions

def in_envelope(state, mu, cov_inv, d_max):
    d = mahalanobis(state, mu, cov_inv)
    return d <= d_max          # else: model output is extrapolation

# On breach: model.authority = 0 ; fall to certified fallback ;
# log operating point as an active-learning candidate for L0

Regime breaches are gold for active learning: they mark exactly the operating points where the model is weakest and new data is most valuable. The breach both protects the machine now and schedules the L0 work that will extend the envelope later, closing the continual-learning loop.