What are you saying precisely? It’s well known that LLMs have non-deterministic output (Ilya Sutskever even claims as such). Are you saying the way it goes about retrieving tokens as deterministic?