But it’s still possible that some M could dominate all semicomputable SWBIs in the weaker sense of losing less than epsilon per step, for all epsilon > 0.
You may already know that, but for the benefit of onlookers: multiplicative weights over F and -F is computable and kinda sorta dominates both F and -F in the weaker sense that you specified. You can make epsilon arbitrarily close to 0 by choosing a multiplier close to 1, at the cost of also incurring an additive constant loss that goes to infinity as epsilon->0. Multiplicative weights over all lower-semicomputable SWBIs dominates them all in the same sense, but it doesn’t seem to be lower-semicomputable itself, only approximable (or maybe I just failed to find a proof).
You may already know that, but for the benefit of onlookers: multiplicative weights over F and -F is computable and kinda sorta dominates both F and -F in the weaker sense that you specified. You can make epsilon arbitrarily close to 0 by choosing a multiplier close to 1, at the cost of also incurring an additive constant loss that goes to infinity as epsilon->0. Multiplicative weights over all lower-semicomputable SWBIs dominates them all in the same sense, but it doesn’t seem to be lower-semicomputable itself, only approximable (or maybe I just failed to find a proof).