Concept Drift Detection

Concept-drift monitors detect when the relationship a model learned between inputs and outputs stops holding, even if the inputs themselves look familiar.

When the mapping changes, not the inputs

Covariate shift is a change in what inputs arrive; concept drift is a change in what those inputs mean. A breeder disruption-precursor model may see familiar-looking magnetics yet the precursor-to-disruption mapping shifts as first-wall conditioning changes. A burner plug model's inputs may look nominal while the potential-to-confinement relationship moves as coil performance ages. Concept drift is the more dangerous of the two because the inputs give no warning.

Detecting it requires labels, which arrive late: the outcome of a pulse is known only after the pulse. Kronos therefore runs concept-drift detection as a backward-looking process on the L0 archive, comparing rolling model error against a stable baseline error rather than comparing input distributions.

Methods

• DDM / EDDM style error-rate trackers with warning and drift thresholds
• ADWIN adaptive windowing over the residual stream
• Page-Hinkley test on cumulative prediction error
• Sliding-window AUC / calibration decay for classifiers

class PageHinkley:
    def __init__(self, delta=0.005, lam=50):
        self.delta, self.lam = delta, lam
        self.mean = 0.0; self.n = 0; self.mT = 0.0
    def update(self, error):
        self.n += 1
        self.mean += (error - self.mean)/self.n
        self.mT += error - self.mean - self.delta
        return self.mT > self.lam      # True -> concept drift

A confirmed concept-drift event is high priority: it means a deployed model's core assumption is stale. It triggers immediate cross-checking against covariate shift, a confidence downgrade in the twin, and a prioritized L0 retrain. If the drifting model holds actuation authority, the rollback path is armed so control can fall back to the last certified artifact.