Pretty fascinating. I think other commenters have uncovered the answer to what is going on under the hood, which is what I came to ask about. Obviously the model can only respond with tokens of some kind. You closed off all the possibilities in terms of other outputs, so it responded with the only remaining possibility: non printing tokens. That is impressive, but not fundamentally different from its using Chinese characters, another creative move. The remaining unanswered question is whether the answers it gave to your consciousness questions were fundamentally different from those it would have given if you’d allowed it the use of, say, yes and no only. That’s an easy experiment which I’ll try myself…