What disruption labels reward: how 11,471 MAST plasmas ended, and what predictors learn from each label set

Open data from the MAST spherical tokamak (FAIR-MAST) are increasingly used to train machine-learning disruption predictors, often with disruption labels taken from the plasma current alone. We classify how each of 11,471 MAST discharges ended, from the programmed plasma current and the sweep rate of the central solenoid at the current quench. Of the 10,616 shots without an operator disruption note, 37.3% (95% CI 36.4-38.2%) were lost during the programmed flat-top with solenoid flux still available, like 72.9% of the noted disruptions; conversely, 45.8% of all current quenches occurred after the solenoid flux was exhausted or during a programmed ramp-down. The result is stable across campaigns (32-42%), classification thresholds (31-43%) and shots outside the sample the rules were set on, and an earlier MAST disruption database built on the same physical criteria agrees with our solenoid classification on 90% of the 6,501 shots both cover. Independently, the operators' own post-shot comments describe a vertical displacement event, a locked mode or an early termination for 10.3% of these un-noted flat-top losses, as often as for noted disruptions (11.2%) and more than three times as often as for flux-exhausted endings (3.0%). Training on each label set shows what it rewards. A gradient-boosting predictor trained on current-only labels reaches a window-level PR-AUC eight times higher than one trained on operator-confirmed labels (0.60 vs 0.07), yet over 50 repeated splits it warns of operator-recorded disruptions no more often than it raises false alarms (-1.4 percentage points, 95% range -3.3 to +0.6), while it warns of 23.8% of flux-exhausted endings. With operator-confirmed labels and magnetic, density and equilibrium signals, the predictor beats random alarms by 8.7 points (95% range 5.4-12.1) at a 5% false-alarm rate over 50 splits, and a predictor that never saw the un-noted flat-top losses alarms on them almost as often as on recorded disruptions. Leaving these doubtful shots out of training raises the share of clear disruptions caught by 9.4 points, a result confirmed on 30 further splits whose test was specified before they ran. All ranges are corrected for overlapping splits and verdicts for multiple testing; code and the per-shot table are public at https://github.com/fazil-firdous/mast-disruption-predictor.