Georg, please forgive me. I thought a sample size of 3 was much too small to make a claim about the typical case, but it appears after doing a computer experiment that I was wrong:

#!/usr/bin/env python
from __future__ import division
import timeit
import random

repeat=100
num=100

def test_func():
    l=1
    for idx in range(10000):
        l=l*idx

timer=timeit.Timer('test_func()','from __main__ import test_func')

data=timer.repeat(repeat=repeat,number=num)

def test_timer():
    sample=random.sample(data,3)
    minval=min(sample)
    onerun=random.choice(data)
    return 1 if minval<onerun else 0

successes=[test_timer() for idx in range(repeat)]
print "Those runs for which the claim in the documentation is true:",successes
s=sum(successes)
l=len(successes)
print "probability that the claim is true: %s/%s = %s"%(s,l,s/l)

Returns:

% timeit-statistics.py
Those runs for which the claim in the documentation is true: [0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1]
probability that the claim is true: 74/100 = 0.74

I'm satisfied the documentation is correct, and I'm really sorry for wasting your time.