Number of severe infections in the experimental group vs number of severe infections in the control/placebo group.

It's past infections. Humanity and the virus are assumed not to have evolved very much in the past year.

You give fifty thousand people injections, 25k get the vaccine and 25k get placebo, but noone knows, because there are two layers of indirection between the person's name and the actual content of the vial. Then you wait until 250 people have fallen ill. If 247 of the 250 were people who got the placebo and 3 got the vaccine, then you assume that actually 247 vaccine recipients would have fallen ill, but only three did, therefore the vaccine protected 244, for an efficiency of 244/247 = 98.7%.

All numbers pulled from a virtual hat, of course.

I believe the trials show zero severe cases after taking the vaccine.