I’m the author. This is a bounded experiment, not a claim of universal model failure. The result here was that when a rule had to hold across multiple mutation paths, code-only prompting repeatedly fixed the named path and missed the equivalent unnamed one; spec-first runs updated all affected surfaces. The page includes limitations and linked artifacts.