Description

neutrinoceros

opened on Mar 28, 2026

Issue body actions

Found this while trying out astropy's test suite against Python 3.15
(xref: astropy/astropy#19503)

>>> import numpy as np
>>> (1.,) + np.float64(2.) # just to demonstrate that this works
array([3.])
>>> from typing import NamedTuple
>>> class T(NamedTuple):
...     x: float
... 
>>> T(x=1.) + np.float64(2.)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
    T(x=1.) + np.float64(2.)
    ~~~~~~~~^~~~~~~~~~~~~~~~
TypeError: can only concatenate tuple (not "numpy.float64") to tuple

In Python 3.14 I get the same output from with a builtin tuple or a NamedTuple.
It feels (to me) a bit more likely to be a regression in CPython, but I expect that the first question I'll get if I report there first is what numpy
devs think about it, so here I am, asking here first.

Activity

1. jorenham commented on Mar 28, 2026

   

   jorenham

   on Mar 28, 2026

   Member

   More actions

   From a type-theoretical perspective this can indeed be considered a bug in CPython. This named tuple called T here is a subtype of tuple[float], and should therefore support at least every operation that a tuple[float] supports. Put differently; this violates Barbara Liskov's substitution principle (wiki).

   Reacted by Clément Robert
2. neutrinoceros commented on Mar 29, 2026

   

   neutrinoceros

   on Mar 29, 2026

   ContributorAuthor

   More actions

   Thanks ! I managed to remove numpy entirely from the reproducer so I think we can just close this one. FTR the report is now at python/cpython#146592

   Reacted by Joren Hammudoglu
3. neutrinoceros

   closed this as completedon Mar 29, 2026

Issue actions