Each hypothesis is a testable claim generated by the seeder and classified into a two-level taxonomy of rank factors with assigned prior probabilities, defining exactly what structural, stylistic, or substantive change might drive model preference.