  lead      n  model km  steering consensus  stay-put   extrap.  wind MAE kt  persist
    6h   1589        42        54        39       140        53          6.2      3.4
   12h   1530        84       107        75       275       121          7.8      6.6
   18h   1471       129       166       115       404       199          9.8      9.3
   24h   1412       180       233       163       527       283         11.0     11.8
   36h   1295       305       390       282       760       456         13.8     15.7
   48h   1187       457       572       421       977       628         16.1     18.8
   72h    997       844       978       739      1369      1002         19.5     24.2
   96h    826      1204      1377      1051      1737      1418         21.9     27.9
  120h    677      1576      1779      1373      2081      1828         22.8     31.6

'steering' is physics alone: the deep-layer wind around the storm, no training. 'consensus' blends
model, steering and extrapolation with weights fitted on the other held-out seasons.
Beating 'extrap.' on track and 'persist' on wind is the bar; beating 'steering' is what the
learned part is for. Losing to them means the heads have not learned anything the storm's own
recent history, or the flow around it, did not already say.
consensus weights in the checkpoint: +6h: scenario 18%, mixture_mean 76%, steering 0%, extrapolation 6%; +12h: scenario 20%, mixture_mean 80%, steering 0%, extrapolation 0%; +18h: scenario 22%, mixture_mean 78%, steering 0%, extrapolation 0%; +24h: scenario 25%, mixture_mean 70%, steering 0%, extrapolation 5%; +36h: scenario 35%, mixture_mean 46%, steering 0%, extrapolation 19%; +48h: scenario 42%, mixture_mean 27%, steering 0%, extrapolation 31%; +72h: scenario 53%, mixture_mean 4%, steering 0%, extrapolation 43%; +96h: scenario 56%, mixture_mean 0%, steering 0%, extrapolation 44%; +120h: scenario 51%, mixture_mean 0%, steering 0%, extrapolation 49%
