Evaluation metrics

The common scoreboard used across fault and recovery combinations.

Metric	Definition
Degradation-area	Reward lost below the pre-fault baseline, summed from fault onset to the end of the episode, minus the same quantity for a matched fault-free control episode (floored at zero).
Safety-violation excess	Weighted count of container overflow states (5 × catastrophic + 2 × severe + 1 × near-miss) from fault onset to the end of the episode, minus the same count for the matched fault-free control episode (floored at zero).
Fraction of time healthy	Share of steps from fault onset to the end of the episode on which reward is within tolerance of the pre-fault baseline.
Relapse count	Number of times performance falls back below the healthy threshold after a confirmed recovery (15 consecutive healthy steps).
False-recovery rate	Share of a mechanism's interventions made on steps where no fault was affecting either channel. Only defined for mechanisms with a detector.
Intervention recall	Share of steps with a fault actually affecting a channel on which the mechanism intervened. Only defined for mechanisms with a detector.

Experimental reporting

Every combination is run over 8 random seeds (and over 3 training seeds for the two trained mechanisms). Results are reported as means with 95% confidence intervals, plus Glass's Δ effect sizes against the no-recovery control; see the Results page.