I cant remember Thompson's stuff: Im talking the general problem. You're correct in that my first step shifts the problem: does it on purpose. Here's what subversion resistant development takes: modular software with sensible interfaces: ability to understand code for human review (closer to algorithm the better); ability to understand compiler passes in isolation; ability to implement toolchain in language of choosing. There are existing flows like this as I illustrated. So, you use them and leverage diverse audience to check results. I mean, would you rather implement CompCert passes by hand or GCC even without optimizations? See the difference? ;)

Now, I did have a method to solve problem you're addressing. You implement an assembler first. Then a macro assembler with macro's for HLL primitives. You can use that immediately to jmplement certified compiler. Alternately, you can pick up Oberon report or Scheme book to implement that to get a true HLL plus compiler. Then you implememt the certified compiler with it. Comprehension, code complexity, and trust are kept manageable by building layer by layer. This, for productivity not security, is how Wirth and Carl first built Lilith then Oberon. Same method will work again and good that ML/Scheme/Oberon folks already gave us doc's plus code to use. Lets use them.