Dear Editor,

first, we like to thank the referees to look at our manuscript. We note that
they both do not find any specific aspects of the manuscript they see a need
for changes to.

While one of the referees feels this as one of the important papers on
device-independent randomness generation this year, the second referee
questions that, in comparison with work of Liu et al. (our reference 10), our
work does not warrant sufficient breakthrough elements to warrant publication
in PRL.

We like to contest that assessment of the second referee for several reasons.
First, the critique seems to be based on a numerical comparison between
obtained key rates, and transformed into a number game. We feel that our work
is not just a number game, but adds a significantly new idea in how to define
what the measurement runs are - that allowed us to operate with a cw source,
and all simplifies experiments based on loophole-free tests based on down
conversion, as it does away with a technical definition of what a round
comprises. 
We also use a randomness extractor [our refs 9, 26] that does not require the
injection of large amount of external randomness, and is thus able to actually
grow the randomness through our experiment. 
Both aspects are  genuinely new in our work, and not found in [10].

Secondly, if one compares numbers, the comparison should be made in a fair way,
which we feel the referee might have missed. The extracted randomness rate of
240 bit per second comes from the short run (43 minutes total time, 26 minutes
data acquisition) with a quantum-safe extractor that actually grows the amount
of random numbers (our references 9, 26), and not a  simple Toeplitz extractor
that requires external randomness added.  I don't believe the data in [10] would
even allow to extract randomness with such an extractor, it certainly was not
done in their experiment. Also, if we reduce our conservative estimation of
the time to the actual data acquisition time of 26 minutes, this number would
be 396 bits per second. As it is not clear from [10] what time was considered,
we can not make a comparison; the difference between acquisition time and
total time in our case is dominated by the random re-orientation of motorized
wave plates. In any case, we still believe that 26 minutes or 43 minutes is
significantly shorter than 111 hours to get any key - definitely more than a
marginal improvement, if one would go for a number game.

Due to the nature of the extractor we chose, we were not able to extract the
randomness in a sensible amount of time for our long run (it would require 120
days of processing). Therefore,  we skipped, as it adds little physics to our
work. If we used the Toeplitz matrix extractors that came in use in [10], the
data rates would have been 943 bits per second of acquisition time, in a
significantly shorter data acquisition time.

So we feel that even if one were to make a number game for publications in
PRL, our scheme would compare very favorably, and pose a significant
advancement. But we believe a manuscript in PRL should not be about this.

We therefore hope that you reconsider the decision of publication of our
manuscript in PRL.

With Best Regards on behalf of all authors,

Christian Kurtsiefer