We're fooling ourselves if we think 'writing a spec' will change
provider's behavior. If you want to provide enticements, why not just
implement some good statistics gathering tools that allow periodically
turning off caching, and analyzing projections from the data so
gathered? It's a lot simpler than all of this fol-de-rol in the proxy
caches.
There's other kinds of sampling, too. For example, you could
selectively turn off caching for randomly assigned host addresses
e.g., hash the host name, and day k of a n-day cycle, turn off caching
for hosts whose hashed IP address mod n == k.