The devil is in the details

People say use smart models to plan and cheap models to implement... but aren't all the devils, well, in the details?

My hypothesis is that cheap models

1. make innocuous small mistakes that compound
2. don't choose the most elegant solution, thus making broken windows, making other coding agents take cut corners (the sloppy way)
3. it takes much more inference later on to both identify and fix it

So I'm trying to prove this - ..