You're wrong. :-) Eviction is managed per node. A write operation on server A will affect the eviction calculation on B (since the write replicates to B). But a read on A does not affect the calculation on B, since B knows nothing about a read on A -- there's no message sent around the cluster saying "A just read /x" as that would add too much overhead.
Nuts, I hate when that happens. :-)
Thanks for the prompt response, Brian.