Yep, that works - thanks. That’s what I have been doing (except using constant/1 instead of fixed_map/1).
I guess I was wondering if it’s advantageous to have a generator which streams over all the inputs and can provide bad inputs for each attribute one-at-a-time, then shrinks to just the ones which fail. But that might cause a combinatorial explosion of potential cases which isn’t feasible to test.
Anyway, your solution is good and works fine. Thanks 
