I second John's point.  In order to capture Java's semantics "on the
nose", you need to be able to model a slot of the T, which can contain
a "null".

Doing that by transforming Java's "T" to Caml's "T option", while
using a more-efficient representation under-the-covers, preserves the
possibility of disabling that efficiency-hack and debugging in a
cleaner environment.

Moreover, Java's semantics for variable-access is that fetching a null
value doesn't raise an exception -- only invoking a method on it
raises an exception.

So one can't use this "initial value" to model Java's "null".

Now, nobody claimed that Java's "null" was the target here, but, hey,
that's my $0.02.

--chet--