This paper gives a very nice overview of the impact of assimilating additional GNSS-RO data and shows logically and systematically the improvements ROMEX data has on the forecast scores using the JEDI-GFS system. First, the impact on short range forecasts are shown by comparing short range forecasts to observations. Secondly, the impact of short and medium range forecast scores are shown with using the IFS as a reference. It is impressive to see that 20 K and full ROMEX show systematically an improvement to e.g. humidity and wind in the short-range; which is consistent to what observations are seeing. Also, for longer lead times the impact is quite substantial by the additional assimilation of GNSS-RO data. However, also some detrimental impacts are illustrated and well documented in this paper. In general, the paper reads nicely and is well structured. I recommend to accept this publications with some minor revisions; stated below:

General comments

General question: “Which version of processing was used for Yunyao or other RO data in your study?” Maybe mention that somewhere.

Specific Comments

p2, l53: I understand you mention only the commercial data operationally assimilated but maybe also mention the Chinese companies (also used in the ROMEX studies). I am not sure if they are assimilated by CMA nowadays?!

P7, Fig.1: I was wondering what is shown here. The total number of RO profiles over the month of September for every 5x5 box? Or the number of average daily profiles over that month?

Also, I think it would be better to use a radially symmetric kernel to estimate the number density for the following reason. If one compares the number for 5x5 lat/lon boxes the area covered over the Tropics is much bigger (and more chances to have RO data) than over the Poles. This would give a wrong impression of where the most data is located.

P9, l221-223: This is true but one has to admit that the horizontal location of that observation point can be different to the tangent point horizontal location- hence, we have ROPP 2D. Maybe mention that.

P11, l252: Mention that September 2022 this is not the full ROMEX period

P13, l323: Maybe indicate the magnitude of this cooling/drying for ROMEX.

Technical comments

Throughout the manuscript correct the spelling of “centre” in ECMWF.

p1, l12: add “daily” after “35,000” and before “RO profiles”

p2, l30: Replace “RO” with “GNSS-RO”. RO is just the way it is measured - it doesn't fit the remainder of this sentence, as it describes GNSS-RO.

p2, l42: Again I would use GNSS-RO to be really accurate but of course you could mention in the text that  with RO data you mean GNSS-RO data. (also l.53)

P4, l94: Change “improvement” to “impact” or “change”

P5, l141: add “daily” after “35,000” and before “profiles”. This keeps coming up at more occasions throughout the manuscript when 35000 profiles are mentioned. Please check.

Summary

This manuscript evaluates the impact of RO observations from ROMEX on the NWP. Using the GFS model and JEDI assimilation system, results show that assimilating additional RO data can further improve the global weather forecasts, with beneficial impacts lasting up to five days. The result also indicates that there is no saturation for the forecast improvement. The manuscirpt is well written, and the experimental design and results are solid. The manuscript could be a valuable contribution to the communities of RO, observations, assimilation, models, and forecasts. I have several comments as below.

1. Section 2 and Figure 1, it is unclear about the differences among ROMEX, ROMEX20K, ROMEX sub dataset. Based on the words, I would guess Fig. 1a + Fig. 1b = Fig. 1c, and Fig. 1a + Fig. 1d = Fig. 1e, while Fig. 1d is not included in Fig. 1b. But based on Fig. 1, it is hard to imagine Fig. 1a + Fig. 1d = Fig. 1e. It would be nice to clearly describe the dataset, which could be consistent with the assimilation experiments.
2. ‘Quality control’ is defined as ‘QC’ at line 211, so that QC can be consistently used later on (e.g., line 229, 236, 351…). How to QC the RO observations are discussed in the manuscript. Does it use the 3 times of standard deviations of the observation error, or something else?
3. I have a curious question about the bias introduced by assimilating the RO. Since the observation error at high altitudes is already large (Fig. 2), which implicitly contains somewhat effect of observation error inflation. Is it possible to conduct bias correction to the RO, like the commonly adopted bias correction for the satellite radiances? Are there systematic features for the bias?
4. Based on Fig. 8, the authors state that “Notably, there is no clear sign of saturation, as most levels continue to show improvement with increasing numbers of RO profiles. However, the degree of this non-saturation appears to depend on both the variable and vertical level and could be influenced by the specific data assimilation configuration.” It seems an overstatement for the ‘no clear sign of saturation’, especially for the state variables at high levels (e.g., T at 400 hPa, wind speed at 250 hPa). It would be nice to discuss these results. Are the non-saturation errors due to the data assimilation algorithm, or observation type, or other potential reasons?

Overview

This paper considers the effect of additional GNSS-RO observations from the ROMEX experiment with the JEDI and GFS systems.  Overall it is well-written, provides an interesting summary and deserves to be published.  I have a few questions on the presentation of the paper, which need to be addressed.  Otherwise it is generally in good shape.

Specific comments

L46: It would be helpful to add Samrat et al (https://rmets.onlinelibrary.wiley.com/doi/10.1002/qj.5002) to the list of publications which show the value of GNSS-RO observations.  Since it is a data-denial study it serves to highlight that GNSS-RO is one of the most important observing systems currently available.

L52: I feel that 2,000 daily profiles from other government missions is an under-estimate.  By my counting we have: Metop-B/C (1100 occs), FY-3D/C (1100 occs), Tandem/Terrasar/Grace/PAZ (900 occs, combined), Sentinel-6A (1100 occs).  All that comes to around 4,200 profiles per day, unless my calculations are off.

L138: The Fengyun missions appear to be missing from this list.

L145: I find this wording confusing "the supplementary profiles are reduced".  Perhaps merge this with the previous sentence: "... is 20,000, meaning that ROMEX20K has approximately 12,000 supplementary profiles per day above the BASE experiment."

L162: I would suggest that the colour scale on Figure 1 is unhelpful, since it is pale in the centre of the range rather than shifting smoothly from light to dark.  If possible, please can the authors update this figure?

L247: I assume from the large relative errors above 40km that the authors are also using a minimum threshold for the observation error (3 micro-radians is typical).  Please can you state what is used.

L253: The sentence beginning "All experiments assimilated" is unnecessary, since it is explained in more detail in the following sentence.  Please can you remove / reword this sentence.

L258: It would be helpful (here or later) to discuss the experimental limitations.  The two issues which seem likely to be the largest are the limited experimental period (only one month) and the limited number of other satellite observations used (no hyperspectral IR, geostationary radiances, atmospheric motion vectors, etc.).  It would be good to mention these here and in the summary, as well as any other issues of which the authors are aware.

L290: Since the acronym MAE is widely used to refer to mean absolute error (similar to the RMSE), the use of MAER could cause some confusion.  Perhaps mean absolute bias reduction (MABR) would be a preferable name.

L302: Since Figure 2 goes up to 55 km, I'm surprised that Figure 3 stops at 40 km.  Does this imply that the observations are only assimilated to this level (which would need stating if true)?  Please could the authors amend the figure, or clarify the assimilation limits?

L379: It is confusing that the figure caption refers to (a) and (c) before (b).  Perhaps the individual plots should be reordered so that wind speed appears as Figure 8(b) so that they are in order.

L612: Whilst Figure 16 demonstrates a degradation in the standard deviation of forecast error above 50 hPa, the authors speak about sources of the biases.  In fact, the changes noted in the presentations from ECMWF and the Met Office largely focused on changes in the forecast bias, rather than the random component of the forecast error.  Therefore, it would be helpful to show plots illustrating the change in forecast bias.  Additionally, those presentations largely discussed changes in the geopotential height bias in the troposphere, whereas this appears to be a degradation in the stratosphere.  It would be helpful for the authors to discuss this difference.