`np.empty` was producing values at the edge of numerical zero.  These values
would apparently cause `jax` to truncate something during and intermediate step
of a comparison operation like `array > 0`, which would produce a result
different from standard NumPy's.