All of Nate Thomas's Comments + Replies

Note that it's unsurprising that a different model categorizes this correctly because the failure was generated from an attack on the particular model we were working with. The relevant question is "given a model, how easy is it to find a failure by attacking that model using our rewriting tools?"