Trying to breed an algorithm to break a substitution cipher.

I'm planning to use three stages:

1. Frequency attack: Breed some genomes for which the frequency distribution is close to that of English
2. Dictionary attack: Make sure all/ most of the words are English
3. Make sure all letters are accounted for: Make sure the genome contains all of the letters A - Z.

Breeding program is written Stage 1 is written, but it seems to converge to a local maximum, so I'm not sure how to get a diverse group of genomes.