If a class has a __numpy_ufunc__ method defined, this disables the usual Python binary operation mechanism, because Numpy's __add__ etc. are implemented via Ufuncs.

Example:

import numpy as np

class Foo(object):
    def __numpy_ufunc__(self, *args, **kw):
        return NotImplemented
    def __add__(self, other):
        return 1
    def __radd__(self, other):
        return 1

x = np.array([1,2,3])
y = Foo()
x + y   

# Traceback (most recent call last):
# TypeError: __numpy_ufunc__ not implemented for this type.

The fix is probably to alter the methods in Numpy's multiarray/number.c so that either (i) the TypeError is caught and converted to NotImplemented, or (ii) a special internal kwarg is passed to the Numpy ufuncs that instructs them to return NotImplemented rather than raising errors.

The problem with (i) is that also other TypeErrors could end up being caught, which may be confusing. (ii) doesn't have this issue, and would allow fixing the issue that ufuncs return NotImplemented in some cases (they shouldn't --- NotImplemented should ever appear only when dealing with Python's binop mechanism).