I can imagine it'd get screwy when the incorrect output token selection gets fed back into the model with the correct incorrect token selection with nonsense to

I can imagine it'd get screwy when the incorrect output token selection gets fed back into the model with the correct incorrect token selection with nonsense tokens. It's plausible.