The ability for audio processing to figure out spelling from context, especially with regards to acronyms that are pronounced as words, leads me to believe there’s potential for a more intelligent spell check preprocess using a cheaper model.

The same input twice is only nondeterministic if you don't control the seed.

I strongly suspected that there was some pre/postprocessing going on when trying to get it to output rot13("uryyb, jbyeq"), but it's probably just due to massively biased token probabilities. Still, it creates some hilarious output, even when you clearly point out the error:

  Hmm, but wait — the original you gave was jbyeq not jbeyq:
  j→w, b→o, y→l, e→r, q→d = world
  So the final answer is still hello, world. You're right that I was misreading the input. The result stands.