Thank you for your revised manuscript. I am sorry for the delay, which was caused by the previous topical editor no longer being available.

I think that you have answered most of the reviewers concerns, but there are still some problems with the manuscript.

Language...

The response to the reviewers and the revised manuscript are both very difficult to understand. I think the main problem is incoherence of language rather than incoherence of thought, but I cannot accept this manuscript until I have seen a version with much improved English.

Some of the problems:
Words are spelled incorrectly;
Words are used inappropriately;
Tenses are wrong;
Singular and plural are mixed up;
Punctuation is missing;
It is often not clear what "it" refers to;
Other grammatical errors that render sentences incomprehensible;
Several typos.

I will send you my copy of your response in which I have highlighted all the parts that I had to read multiple times in order to understand your meaning. Doing this helped me to understand your work. I hope it will also help you to improve the manuscript.

Note that "similar" means "resembling without being identical". If you write that something is similar to something else then you also need to mention in which ways it is different. If you actually mean identical/the same, then say so.
For example,
"Amongst these parameters, along with the basal effective pressure, the large scale bedrock curvature and/or sub-grid roughness could be used, similarly to Briggs et al. (2013). "
could become,
"Amongst these parameters, along with the basal effective pressure, the large scale bedrock curvature and/or sub-grid roughness could be used, as in Briggs et al. (2013).",
should this be the meaning that you intend.

Note, also, RMSE is not yielded or presented.
"... have an RMSE of ...", is sufficient.

On to the science...

Section 3 "Calibration of the Antarctic ice sheet"

This section is unconvincing. I think an additional paragraph is needed here, to explain what you are doing and why. What I would expect to see for a model calibration is the development of a cost function that specifies a PDF of what you believe to be plausible models. The final set of models is then sampled from this PDF. In order to develop this PDF, discussion of the observations, the observational error, and the model adequacy (or discrepancy) need to be included.

What you have actually done seems rather bizarre. You appear to have 3 observations (total volume, ice thickness, ice velocity), and yet you have used only one of these (thickness). You have included in your final ensemble only those models with the lowest RMSE compared to that single metric. Observational error is not mentioned. Discussion of model (or in this case experimental) adequacy suddenly appears in a paragraph of the Discussion (the one that starts "Although widely used... ") where you explain that you do not anyway believe the result of your calibration. I conclude from this that you in no way believe that your 12 "best" ensemble members are in fact your "best"! Would you be planning to use this set of 12 in any future work?

I might be being stupid, but I cannot find figures showing the observed thickness, or uncertainty in that thickness. In addition the plots of RMSE are not very informative as they do not indicate whether the ice is too thin or too thick. I also cannot tell whether (and where) the whole ensemble is biased, which is surely important information.

I wonder whether what you are doing is more of a sensitivity test than a calibration. It seems to me that what you are investigating is the range of model behaviours after a strong constraint is placed on ice thickness. If so, then the section needs quite a bit of rewriting to make this clear.

Numerical solution methods...

I have rarely seen a manuscript in which the reviewers clamour so loudly for the code to be made available. While some description of the numerical solution methods has been added to the manuscript, it is still rather generic (eg "an upwind scheme" !!) and I think this is insufficient for GMD's aim of "scientific reproducibility". One possibility would be writing the difference schemes in an appendix or in the supplement, but this is unlikely to fully satisfy all readers, and a simpler solution would be to be more specific in the manuscript while also making the code developed in this paper available in the supplement.

I do not fully understand the grounds under which the code is (or is not) being made available at present. You wrote in your response,
"The public distribution needs to be done with a specific license (e.g. GNU public licence). Such a license is currently lacking in GRISLI. The model has been developed from the 90ies and has benefited from the additions from numerous contributors. To put a license to the current model, we have to get the permissions to all the contributors from the past 20 years. To date, we do not have this permission from all the contributors. However, as stated in the manuscript, any potential users are encouraged to get in touch with C. Dumas, A. Quiquet or C. Ritz, to start a collaboration."

My understanding is that it is the employers, rather than the individuals, who need to give permission. How many of these employers have you been unable to track down (are some no longer active?), and which parts of the code are still restricted? If you upload all parts of the code that are not restricted, will this include the parts of the code that the reviewers want to see (ie the numerical solution to the equations outlined in this manuscript). If so then this may be satisfactory solution for the present manuscript. It seems strange that you can grant access upon request, but not release the code - what is the current license that users get when given access to the repository? I do understand that it can take a long time (years!) to achieve full code access for older models, but I would like to see some evidence in the manuscript that this is being actively pursued. It is already foreseeable that in the future all code developed in scientific publications will have to be made publicly available, so you are strongly encouraged to make this a reality for your own model sooner rather than later.

Thank you for the revised manuscript. It is amazing how much a manuscript improves when even only a few of the worst language mistakes are corrected (and you have done much more than such a minimum!). There is an obvious typo in the abstract, but I understand the manuscript well enough now to leave the rest to the copy editors (their English is much better than mine anyway).

The outstanding problem is the calibration. I asked for an introductory paragraph explaining the purpose of this, and I still think it is required because I do not know what the point of the exercise is. If the aim is to calibrate the model why run on 12 members? If the aim is to look at uncertainty, why run the 12 that are closest to the data? The data look like they have fairly high uncertainty - what range of RMSE might you expect from a perfect model? Are the 12 ensemble members all within or outside of that range? Usually, when looking at uncertainty, the aim is to explore some range of plausible models. There is always quite a bit of subjectivity involved, but something people might do is loosely re-weight the ensemble by the data (including effects of model and data error) and then resample from that re-weighted ensemble. The reason it is important to explain this much more clearly is so that your results (and the ensemble members themselves) may be used appropriately in future. Also, without this guidance the reader does not know what conclusion to draw from Figure 14.

I think that the code availability is now sufficient to comply with current GMD requirements. Thank you for addressing that.

There has not however been much success in answering my questions about the calibration apart from calling the ensemble member "plausible", without giving any reasons why you think this is the case. I think you might have misunderstood what I am asking. I am not asking why you do not use transient experiments for your calibration, neither am I asking for more details on what was actually done. Rather, it is necessary for you to explain the basis for the calibration experiment and what the resulting ensemble means and therefore what it is useful for. I still think that an additional paragraph to this end at the start of section 3 is required - it is confusing to jump straight from the calibration subtitle into the method subtitle without explaining why you are doing it. I am sure this stuff is inside the heads of you and your co-authors, but readers are not telepathic - I just want it written down so that readers can understand the purpose of your work!

From last time,
"I do not know what the point of the exercise is. If the aim is to calibrate the model why run on 12 members? If the aim is to look at uncertainty, why run the 12 that are closest to the data? The data look like they have fairly high uncertainty - what range of RMSE might you expect from a perfect model? Are the 12 ensemble members all within or outside of that range? Usually, when looking at uncertainty, the aim is to explore some range of plausible models. There is always quite a bit of subjectivity involved, but something people might do is loosely re-weight the ensemble by the data (including effects of model and data error) and then resample from that re-weighted ensemble. The reason it is important to explain this much more clearly is so that your results (and the ensemble members themselves) may be used appropriately in future. Also, without this guidance the reader does not know what conclusion to draw from Figure 14."

I previously suggested that perhaps your calibration+transient ensemble experiment was closer to being a sensitivity test, as it shows that model versions with similar fit to the data can have different parameters and quite different results. So now I am now very confused, because your latest Author's response appears to support this line of reasoning, but the text of the manuscript does not! Maybe I have missed a paragraph in the manuscript (in which case I am sorry and please point it out to me) but I can't see where you include these points from your own response,
"However, using these 24 plausible models can help the reader to grasp the GRISLI result spread (Fig. 14) for models yielding a relatively similar present-day ice sheet. We acknowledge that the choice of 12x2 ensemble members is arbitrary and this number is too low to infer statistically meaningful results in terms of model uncertainty. However, even with our relatively coarse resolution, a 400 kyr-simulation represent a nonnegligible computing time that has to be added to the 600 ensemble members of Sec. 3. As a result, Fig. 14 aims at illustrating the spread in ice volume within the 12x2 ensemble members calibrated to reproduce the present-day Antarctic ice thickness."

Note - comments on personal computing cost are entirely appropriate in GMD manuscripts, as this effectively constrains the state of the art that is possible for individual scientists. Such comments can also be useful for helping authors understand the methods adopted. It actually quite often happens that reviewers unwittingly suggest additional experiments that are unfortunately just not presently possible for the authors to complete. But I am not suggesting that you redo your experiment better this time around - I only want you to explain it better and perhaps point the way for future researchers.

I am not sure why this took so many iterations, and my feeling is that if you had taken longer to carefully think about it from a reader's point of view, before bouncing it back to me, then it would have taken less time in total. This version is now sufficiently well explained for the manuscript to be comprehended by at least other earth system modellers, if not the whole of the geoscience community, so I am happy to accept the paper. Congratulations!