Lever #1 from the report's ยง8 Next list, run for real: at every
walk-forward fold, the champion the report already shows was picked purely by
re-scoring on the full training span at the end (the exact objective being
maximized -- zero independent check). Here, each fold's 18-month training span
is split 75% train-sub / 25% validation, chronologically, with
the scaler fit on train-sub only. Every generation, the genome that's best on
training is also scored, read-only, on the validation slice -- one
look per generation, never fed back into breeding. Whichever generation's
snapshot scored best on validation across the whole run becomes the
"validation-selected" champion. Both that champion AND the original
training-selected one are then scored on the fold's real, never-touched test
window, so the two methods are directly comparable on the same out-of-sample
data. See the discussion of why this doesn't fully escape the same
multiple-comparisons risk in the chat that requested this page -- validation
selection reduces bias, it doesn't eliminate it.