I wonder if anyone's tested the reverse, whether Claude/GPT models produce worse code for an ostensible Chinese or Russian government agency or contractor.