Smallest budget: fewest candidates, smallest validation set, lowest trial cap.
upper bound on optimization trials for this mode.
number of instruction/demo candidates generated per module.
number of validation examples used to score candidates.