Sub-Threshold Alarming and False-Alarm Control
Detecting precursors below protection thresholds only helps if the alarm rate is low enough to act on; Kronos tunes thresholds against a cost of false vs missed alarms.
The operating-point trade
A precursor detector is defined by its threshold on the ensemble score. Lower it and you catch more precursors earlier but raise false alarms that erode trust and cause needless ramp-downs; raise it and you miss precursors. Kronos sets the operating point explicitly on the detector's characteristic curve, weighting a missed precursor far more heavily than a false alarm, but not infinitely, because chronic false alarms are themselves a safety hazard through desensitization.
- Miss cost: high, a missed precursor forfeits the lead time the ensemble exists to buy
- False-alarm cost: real, needless ramp-downs and operator distrust
- Chosen point: high sensitivity for magnet-quench class; agreement required to fire
- Hysteresis: alarm set and clear thresholds differ to prevent chatter
Two-stage alarming
Kronos separates a sensitive watch stage from an act stage. The watch stage flags a candidate precursor at a low threshold and raises attention (tightening the twin's monitoring, requesting corroborating diagnostics). Only if corroboration accumulates, multiple detectors agreeing, the forecast-residual diverging, a physically consistent localization, does the act stage request MPC to move to a safer point. This staged design gets early warning without early over-reaction.
Thresholds are per-failure-class and context-dependent: the acceptable false-alarm rate during a routine breeder flat-top differs from that during a commissioning ramp. The twin supplies context (operating phase, current, field) so thresholds adapt rather than being globally fixed.
None of this touches the hard protection threshold. Sub-threshold alarming lives entirely below the L1 hardware failsafe trip level; its whole purpose is to reduce how often that hard trip is ever reached, by acting earlier and more gently.