We would like to thank the referee for providing feedback on our manuscript. We give our response to the referee's comments and we explain the changes made to the manuscript below.

>> The article is well-formatted with suitable length. I recommend this article to be published in PRD.

We thank the referee for the positive recommendation.

>> However, I suggest to revise the article so as to clarify several points.
>> At the beginning of page 4, it would be better to describe the
actual number (or plot) of the phase delays and time delays.

We agree with the referee and we have included the time delay from the filters in the text on page 4. We changed the sentence ``Thus, injections have known, but uncompensated, delays that we take into account during injection recovery.'' to ``These delays, on the order of 240~$\mu$s, are taken into account during injection recovery.''

>> Fig. 2 shows a comparison of the injection capability and the actual
signal amplitude. For that purpose, ~1kHz cutoff (the 'second' effect
described in the text) should be also included in the plot.

We made the changes requested by updating Fig. 2 and  we included a dashed line beyond 1kHz to indicate that the signal fidelity is limited beyond 1kHz. And we included the following sentence in the caption: ``The dashed blue curve indicates that the fidelity of the induced displacements degrades above 1kHz due to the need to roll off the inverse actuation filters to maintain stability near the Nyquist frequency.''

>> In Section III A, actual parameters of the injected signals, such as
mass, distance and sky position, should be explicitly described.

We agree with the referee and we have included more information in the description of the waveforms.
The paragraph in Section III A was revised to the following: `'`For GW150914 and GW151226, we injected ten waveforms coherently into the two detectors after collecting enough data to confidently establish a detection.
The GW150914 hardware injections were generated with the SEOBNRv2 waveform approximant and included systems with component spins aligned with the angular momentum of the binary~\cite{seobnrv2}.
The GW150914 waveforms had a total mass from [68~M${}_{\odot}$, 79~M${}_{\odot}$] in the source frame, mass ratios from 1 to 1.8, and distances from [250~Mpc, 530~Mpc].
Mass ratio is defined as $m_1 / m_2$ where $m_1 > m_2$.
These signals were injected October 2 to October 6, 2015.
The GW151226 hardware injections were generated with the precessing waveform approximant IMRPhenomPv2~\cite{imrphenom,imrphenom2} and injected on January 11, 2016.
The GW151226 waveforms had a total mass from [25~M${}_{\odot}$, 30~M${}_{\odot}$] in the source frame, mass ratios from 1 to 4.3, and distances from [240~Mpc, 580~Mpc].
For both the GW150914 and GW151226 waveforms the sky positions were selected to be on the same triangulation ring as the corresponding astrophysical event.''

>> How much are the time differences to estimate the values and errors
of the software injection rho in the Fig. 3? Is it sufficiently small
compared to the drift of the detector noise level?

The software injections were separated by 250 seconds. This is much longer than the waveforms that are on the order of a second. The entire set of software injections were injected over a 3.5 hour interval during the same lock stretch as the hardware injection. On these times scales, the state of the detector in observing mode does not change significantly. For each hardware injection, the largest difference between the hardware injection time and the furthest software injection time is between 4 and 8 hours.

For an idea how much the detector noise level changes look at Fig. 4 in [8]. In Fig. 4 our GW150914 injections were done around 3.5 weeks. You can see some variations on the order of percents over a few hour time scale. This change in distance corresponds to a similar change in the recovered signal-to-noise ratio.

We have added the following sentence to the paper: ``Detector data within hours of the hardware injections was selected for adding software injections since the sensitivity of the detectors does not significantly vary on these timescales [8].''

[8] B. P. Abbott et al. (Virgo, LIGO Scientific), Class. Quantum Grav. 33, 134001 (2016).

>> How does one estimate the injected SNR in Fig.7? How much is the error for them? How much are the estimation errors for the reconstructed chirp masses in the right panel of Fig. 7?

The SNR in Fig. 7 is proportional to the "coherent energy" in the gravitational wave signal; see the paper at https://arxiv.org/pdf/0802.3232.pdf. The injected (theoretical) SNR is determined by calculating the coherent energy of the waveform and comparing to the strain noise. This is in contrast to the recovered SNR, which is output by the pipeline without prior knowledge about the waveform. The uncertainty is indicated by error bars added to the plot. For a discussion of the Coherent WaveBurst estimation of chirp mass, please see [49]. Since our paper here is a broad overview of the hardware injection system, we think it is best to avoid going into the detailed inner workings of individual search pipelines such as Coherent WaveBurst. That said, we note that the scatter of points in the right panel of Fig. 7 provides a rough idea of the typical uncertainty (~5 msun for events like GW150914); see also Fig. 4 in [49].

[49] V. Tiwari, S. Klimenko, V. Necula, and G. Mitselmakher, Class. Quantum Grav. 33, 01LT01 (2016).

>> On page 10, how did the authors obtain the error number for the
recovered $\Omega_{gw}$? It would be better to explicitly show the injected
amplitude in Fig. 10.

The uncertainty is calculated using Eq. 5.1 of Allen and Romano [50]. This is the same calculation that has been done in stochastic searches since the early days of Initial LIGO.

The injected amplitude is shown in Fig. 10. It is the red circle.

[50] B. Allen and J. D. Romano, Phys. Rev. D59, 102001