I already wrote http://bugs.python.org/msg270548, but can't seem to reopen the issue, so I think the best thing is to report the bug separately.

So, in issue http://bugs.python.org/issue21136, performance enhancement https://hg.python.org/cpython/rev/91d7fadac271/ enabled a shortcut for some operations (__pow__ among them) to avoid reducing the result to lowest terms if it can be concluded it's already reduced.

However, the logic has a corner case that was handled incorrectly. If a negative Fraction is raised to a negative integer, the result is a Fraction with a negative denominator, which is not normalized and in fact breaks many other operations (which rightly assume their operands to be normalized).

>>> import fractions
>>> fractions.Fraction(-1, 2) ** -1
Fraction(2, -1)
>>> _ == -2
False

Of course, the easy fix would be to change line 52 of fractions.py to _normalize=True - but that would reintroduce the slowness talked about in http://bugs.python.org/issue21136#msg215409. Alternative fix is to branch depending on the sign of the base, which might be messy if we intend to be completely general -- for example, in finite quotient rings, fractions don't have well-defined signs.

[BTW I'm not quite sure why the code in fractions.py is so general, with many weird cases trying to support every imaginable construction - maybe someone really wanted such a level of generality. I don't, so I'd be fine with this working only on ordinary int/int Fractions.]