A Question Asked One Step Earlier
The eviction lesson compared ways of choosing a victim when the cache is full. Caffeine's design is about the question before that one: when a new key arrives and the cache is full, should it be let in at all.
That sounds like a distinction without a difference until a batch job runs.
An LRU has no way to refuse: the newest key is by definition the most recent. So a scan of a million keys evicts a million entries, one at a time.
Read admission and eviction as two separate decisions with two separate policies. Almost every cache has the second and almost none have the first, which is why a nightly export can quietly halve a hit rate.
The Control Comes First
Every policy saw the identical request stream, and the hit rates are exact counts.
Nothing in this chart would make anybody change anything. That is the point of running it.
Run the uninteresting case before the interesting one. A comparison that only ever measured the hard workload cannot tell you whether the winner won because of the thing you are about to claim, or because it was better all along.