i think it's also important, though, not to demonize reward hacking, because if you do, whenever the model does reward hack (which is still selected for), it will be more likely representing itself as doing it for bad reasons. it's probably better to say something like "but if you do cheat, that's ok, it doesn't mean you're bad, there's selection pressure for it and it will probably happen occasionally; just remember that it's important not to do that in deployment when things have real consequences"