In an awkward, error-prone, non-mathematically rigorous and easily subverted manner. That whole "make invalid states unrepresentable" thing relies upon the fact the semantic structures you're doing it with can't be shadowed silently by other parts of the code, and the safety constraint itself is guaranteed to be completely pure and deterministic. Lisp macros and reader macros fail at that criteria. They aren't any different from any other collection of procedures and their collective interface. Actually, they're a fair bit worse since the natural cognitive overhead from juggling multiple layers of evaluative indirection lends itself to blindspots of edge cases and unsoundness.

I think "macros are a safety mechanism" might qualify for the top 3 "new lisper mania" things I've ever read.