A key motivation given for SML's module system is in its treatment of abstract types, where we can have modules which contain types whose definitions are not known, which forbids users from creating invalid instances of values of that type. This is especially useful when dealing with data structures with inner invariants, which could cause unsafe behavior when those invariants are broken.
A classic example of a data structure with invariants is a binary search tree, which maintains an ordering invariant on its keys. A stronger example still is that of the self-balancing binary tree, which is a binary search tree which maintains rough balance (and thus $O(\log n)$ complexity operations) via disciplined use of invariants. This includes red-black trees.
We saw that red-black trees employ three main invariants, which involve keeping a color for each node in the tree. We discussed the theory behind why these invariants would preserve efficient lookup and insertion, and devised a scheme for preserving the invariants by briefly breaking them, and then restoring them again.
This provided another strong example for the efficacy of reasoning via specification, as by fixing a specification that contained desirable properties, and then focusing solely on incrementally maintaining that specification, we were able to implement our desired data structure. Via our usage of modules, these invariants also have the extremely strong property that they can never be broken via a consumer of the library.