LEAKAGE: validation seasons [2017, 2019, 2021] | training seasons 1979 - 2016 (38 seasons) | overlap: none | test seasons 2018/2020/2022 in training: none
weather backbone trained on 2011-2015 ERA5; storm-head training seasons inside that window: [2011, 2012, 2013, 2014, 2015]

CONSISTENCY: landfall 'by this lead' probability falls as lead grows in 20% of forecasts (a cumulative probability must never fall); worst drop 12%
RI: P(bigger jump) exceeds P(smaller jump) in 0% of forecasts (must never happen)
wind radii: a 50/64 kt radius larger than the 34/50 kt one in 21% of forecasts (must never happen)
wind radii: negative radius in 0% of forecasts
two intensity heads disagree: |eyewall peak wind - intensity head| mean 1.5 kt, >20 kt apart in 0% of lead-forecasts
radius of max wind larger than the 34 kt radius in 17% of lead-forecasts (physically impossible)
intensity head second channel (pressure) range: 942.12 to 1023.62
