Why it is a question
THEORY-376 holds what Dar, Franklin, Jónsson, Srivastava and Tan (LIT-1113) simulated: describing cached data by the predicates of past queries matched page caching on a clustered index, beat it on an unclustered one, and on an unindexed attribute with a disk cache answered over 60% of 1% queries without the server. The simulation charged nothing for computing probe and remainder queries, whose formulas need not be convex or connected, or for maintaining region descriptions. The saving it measured is the gross saving.
What would answer it
The condition as it stood, unchanged: a measurement in an implemented client-server system, or a simulation that charges the client for computing probe and remainder queries and for maintaining region descriptions, that reproduces the ordering (semantic caching about equal to page caching under good clustering, better under bad clustering, and faster when queries are answered wholly from cache). A further simulation that leaves reasoning about cache contents free does not answer it.