I think in the last example, it's going straight from arbitrary precision to 54 significant digit, bypassing float64 entirely, hence why it looks different from

I think in the last example, it's going straight from arbitrary precision to 54 significant digit, bypassing float64 entirely, hence why it looks different from the middle example.