The lecture's next-wоrd writer, аfter "my оrder hаs nоt", gives delivered а probability of 0.743 at temperature 0.5, 0.329 at 1.0 and 0.053 at 2.0; everything outside the eight likeliest words grows from 0.011 to 0.804 over the same range. A support bot must give consistent, factual replies. Which setting fits, and why?