1. Imagine a system where the punishment for murder is exactly the same as the punishment for attempted murder, but that punishment has an element of chance: you might get a longer or a shorter sentence.
This doesn't seem particularly unreasonable. Now we're going to make successive modifications to the system, which (Lewis claims) don't obviously make it worse.
2. Now suppose that we attempt to make the punishment for more wholehearted attempts at murder greater. It's hard to tell what was in the criminal's heart, so as a reasonable proxy we try to evaluate, somehow, how likely the murder attempt was to succceed, and we punish more-likely-to-succeed attempts more harshly.
3. Now suppose, more specifically, that the way in which we punish more-likely-to-succeed attempts more harshly is by increasing the probability of getting the longer sentence. In fact, we'll say that the probability of getting the longer sentence is to be the same as our estimate of how likely the attempt at murder was to succeed.
4. How to estimate that probability? One effective but obviously impractical way would be to run some sort of reenactment or simulation of the crime, and give the longer sentence if and only if the victim dies. Aha! But we can simplify and improve this by using the original crime as the reenactment, and give the longer sentence if and only if the actual victim died.
And now we've arrived at pretty much our present practice!