I gather you are referring to "Mutation Testing".  Now I know nothing  
about that topic, but I just read about it on the web, so correct me if I  
am wrong.  My understanding is that you take one program and a suite of  
tests for that program then generate a whole load of minor mutations of  
the program in an attempt to find deficiencies in the test suite (i.e.  
incorrect programs that pass the tests).

I'm not sure exactly what you mean by seed.  I assume you don't mean the  
seed for a random number generator.  It's also not clear to me how you  
would distinguish between a mutant that passes the tests because it is  
functionally-equivalent mutation of the correct program and a buggy mutant  
that passes the tests because the tests suck.  But anyway...

Writing a program to mutate Java source doesn't sound like fun.  Any naive  
approach is going to generate invalid programs more often than not.  This  
is not my area of expertise but I guess you would have to parse the source  
into an abstract syntax tree and mutate that.  Is it sufficient to simply  
re-order a few statements and adjust a few values, or do the mutations  
need to be more elaborate?

There seems to be some overlap here with methods used for genetic  
programming, so it might be worth looking at GP tools.  However, because  
of the problems with generating invalid programs, GP usually uses Lisp or  
domain-specific languages that are easy to mutate.  Languages like Java  
and C++ are generally too much hassle.

Dan.