It could be that small variations in prompting lead to large differences in quality of output, especially over longer horizons.

I’m saying it’s probably multiple factors and both you and GP are right.