I think this affects negation everywhere, also including effect conditions and axioms.

In domains with negated goals, using the option "--relaxed" produces a trivially unsolvable instance. We probably have the same issue with negated preconditions.

We could discuss how more complex cases are handled, e.g. compile away negated goals before the relaxation. Alternatively, we could produce an error when a user asks for the relaxation of a domain that is not in positive normal form. 

I consider the current behavior of silently producing an unsolvable instance as a bug.