This is a review of the paper titled “Model-based evaluation of cloud geometry and droplet size retrievals from 2-D polarized measurements of specMACS” submitted to AMT by Volkmer et al. The paper is well-written and the topic is certainly relevant for AMT. However, the work and results are largely similar to those previously published. As explained below I think the authors can expand the analysis a bit to make this paper a very useful contribution to the field. Below I list a few general comments followed by a few specific comments. I would recommend publication of the manuscript after these points are addressed.

General comments:

1. The results are largely consistent with those of Alexandrov et al. 2012, who used essentially the same retrieval approach and the same 3D radiative transfer model. In that sense the current paper does not add so much to the available literature. I would suggest to add a slightly deeper analysis of the cases where the retrievals substantially deviate from the truth. Are cases with large errors in Reff and Veff related? Do they correspond to low optical thickness cases? Do these retrievals exhibit a bad fit of the simulations to the data? Is there any way such cases could be identified in case of real retrievals?
2. The paper refers to earlier paper for the details about the drop size distribution retrievals. However, I think it is crucial information to understand this paper and suggest that a short description of the retrieval procedure is included.
3. In line 114 it is stated that k = 0.8 is assumed for the calculations of effective radius from the assumed number concentrations. However, in line 116 it is stated that an effective variance (veff) of 0.1 is assumed. This is inconsistent, as k and veff are related by k = (1-veff)*(1-2*veff) for the gamma distributions assumed here (see for example Grosvenor et al. 2020 Eq. 13; https://doi.org/10.1029/2017RG000593). For an assumed veff=0.1, k would be 0.72. I do not think this matters much for the analysis in the paper, so I will not ask to correct this. It should be noted in the paper, however.
4. It is not clear to me if “evolution” of the cloud also includes movement of these clouds. If so, please make this clear and mention the windspeed (profile) assumed in the paper and also indicate how much the clouds generally move during the cloudbow observations.

Specific comments:

Line 133: For clarity, please change to “acquisition time of 8Hz of SpecMACS.”

Line 206: A flight time of 30 seconds is simulated. Please specify what area is covered by all viewing angles in such a time.

Line 252: The accuracy of stereo heights retrievals can be compared to those estimated by Sinclair et al. 2017 (https://doi.org/10.5194/amt-10-2361-2017), who analyzed data with many different cloud types.  

Line 284: If the optical path length for each viewing angle is similar, won’t different viewing angles see slightly different physical vertical locations in the cloud top? Then each angle may see slightly different size distributions. I guess this effect is minor, but it might be good to estimate this. Also, which viewing angle is taken into account for the analysis here?

Line 332: Please give the mean and standard deviation of the differences here. They are only given in the conclusions.

Figure 4, 5 and 6. In the scatterplots all the low values are hard to see. Please change the colorbar.

This is a review of “Model-based evaluation of cloud geometry and droplet size retrievals from 2-D polarized measurements of specMACS” that has been submitted to AMT by Volkmer et al. I concur with the comments given by Reviewer 1 regarding the quality of the writing, and also with the suggestions for improvement.  A few additional suggestions for improving the paper and minor editorial comments are given below.

General Comments:

1. The authors make the claim that they are “validating” retrieval results in several places in the introduction. Given the way they are comparing “model” and “retrieval” results I am not sure that is entirely the case, as there is no evaluation of whether polarized reflectance misfits are related to “errors” in retrieved properties. There is also no evaluation of the effects of noise on the retrievals. Nonetheless the use of a simple model sampling scheme to relate observations to model variables is of considerable value for understanding how to use polarized observations for model evaluation and should probably be more emphasized.
2. The reverse Monte-Carlo sampling of singly scattered photons described at the beginning of Section 5 is the crux of the paper and presents a reasonable and simple way of sampling model output, in order to evaluate it against polarimetric observations. However, the rationale for the use of “the average of all scattering event locations” (line 199) in sampling the model fields and its consequences should be discussed.  For example, a “model” simple average, unweighted by the probability of the path will tend to give a sample that is biased deeper into the cloud than that of the optical signal if singly scattered light dominates. In contrast, if multiply scattered light (e.g. multiple forward scattering events caused by the diffraction peak in the phase function) dominates the signal the “model” simple average will give a sample that is biased higher in the cloud than the optical signal.  This difference between the “model” sampling and what one expects to be the source of the optical signal is one potential explanation for the compensating biases in the stereo cloud top height results.  Some discussion of the single scattering “model” versus full Monte Carlo sampling would therefore be helpful.
3. Figures 4f & l, 5f and l and 6f and l are not useful. The statistics of the regressions are informative but for the reader presenting histograms of differences would be a more effective use of the graphic.
4. As noted in point 2 there will be a difference in the vertical weighting of the “model” sample and that which is expected from the polarimetric retrievals.  It is particularly important to note this when making a comparison between the effective variance retrievals and the model since the effective variance itself is constant.  This means that there is no underlying signal, and the comparison is primarily a comparison of differences in sampling.  An additional point regarding the effective variance comparison is that the retrievals have clearly failed when veff=0.32.  Some additional comment on this and ideally examining whether there are a particular range of effective radii where this failure occurs would be desirable.
5. While I do not think that additional simulations are in order the authors should note that an effective variance of 0.1 is quite large for a cloud top size distribution and this should be born in mind when planning future work.

Editorial Suggestions

Rewrite Eq.(6) as v_eff_tot= veff + (reff_4*reff_2/reff_3^2-1)*(1+veff) where reff_4 is the 4^th moment of the effective radius etc.  I suggest this because it is then clear that sampling variability in reff can only increase the apparent effective variance. 

Where the Marshak et al. (1998) paper is cited at line 121 it should be noted that there conclusions are for overcast clouds. E.g. insert the word overcast between “for” and “marine” on that line.