Detecting Shadow-to-Plant Divergence
When the plant departs from the twin's forecast, that gap is both a model-drift alarm and a physical anomaly signal.
Divergence as a signal
The predictive shadow forecasts the plant; the plant then reveals what actually happened. The difference, the innovation in estimator terms, is information. A small, unbiased innovation means the twin is tracking well. A growing or biased innovation means one of two things: the twin's model has drifted, or the machine is doing something the validated physics did not predict. Both are important, and Kronos monitors the divergence explicitly.
Two interpretations, two responses
- Slow, systematic drift: model is stale -> flag for recalibration/retraining at L0
- Sudden, localized divergence: physical anomaly -> feed the anomaly ensemble
- Whitened, in-bounds innovation: healthy tracking, no action
- Divergence with rising reconstruction/coupling residual: reduce twin confidence
The forecast-residual detector in the anomaly ensemble is exactly this divergence test applied for safety: an event that departs from validated physics is often the earliest and most meaningful precursor, because it does not depend on having seen that failure mode before. A plug-density collapse or a disruption-precursor mode shows up as the plant diverging from the shadow before any single-channel threshold trips.
Kronos whitens the divergence by the estimator's predicted innovation covariance, so the test accounts for how uncertain the forecast was: a large divergence on a quantity the twin already flagged as uncertain is less alarming than a small divergence on one it was confident about. This keeps the test statistically meaningful across the operating space.
Persistent healthy-state divergence is the trigger for the drift-monitoring workflow: the model is scheduled for recalibration or retraining offline, and until then the twin lowers its confidence in the affected quantities. This closes the loop that keeps the twin honest over the machine's life, especially as materials activate and geometry ages under neutron fluence.