The authors dynamically downscale output from the CESM Decadal Prediction Large Ensemble (CESM-DPLE) using a high-resolution regional ocean model, ROMS-DOWN, with a horizontal resolution of ~5 km. The study focuses on the multiyear prediction skill of ROMS-DOWN on the U.S. Northeast Shelf, which complements previous work on seasonal and decadal prediction in this region. Using a set of metrics including the mean state and interannual variability of key physical variables, as well as the representation of important regional phenomena such as the Mid-Atlantic Bight cold pool and slope-water mixing, the authors demonstrate substantial improvements in ROMS-DOWN relative to the CESM-DPLE simulations.

Overall, the manuscript is well written, and the analyses are clear and comprehensive. I believe this study makes a valuable contribution to the literature on regional downscaling and ocean predictability. I have only a few minor comments as listed below.

Minor comments:                                                                                             

1. Figure 2-3: Why does ROMS-DOWN sometimes perform better than ROMS-HIND? I expected the opposite since ROMS-HIND is forced by reanalysis atmospheric anomalies that are constrained by data. Could the authors explain?
2. Line 236: How is NES defined exactly? Is it a combination of GOM, GB, and MAB?
3. Lines 395 and 494: Have the authors looked at the variability of the Gulf Stream position (e.g., the GS index; Ross et al. 2023) or its proxy in ROMS-DOWN? Previous studies have shown that GOM temperature is strongly correlated with the meridional shifts of the GS (Pershing et al., 2015). Therefore, the high prediction skill for NES SST may partly reflect a realistic simulation of GS displacement, as the authors note. I do not think additional analyses are necessary, but a more quantitative comment on the representation of GS displacement in ROMS-DOWN would be helpful.

Technical corrections:

Figure 12: label (h) is repeated. Please revise the plot.

Caption of Figure 14: “Shelf Water (SSW)” -> “Scotian Shelf Water (SSW)”

Figure 14: “label (e) Lead Year 8” -> “label (h) …”

Summary: 

This paper provides a detailed analysis of the performance of a downscaled predicition system for the American Northeast Shelf (NES) region, over the challenging sub-decadal timescale. The paper is well written, and the results will be useful for a range of reader – both model developers and potential users of model output.

However, the paper would benefit from some revisions that would help improve clarity of results for the reader. Firstly, there are a large number of figures that can be hard to compare at times, so some thought on use of tables for key statistics could be useful. I also feel more could be done to clarify whether the model is intended to benefit on-shelf or off-shelf predictions. Some discussion on this for both the motivation and conclusion could help target the analysis and conclusions.

General comments:

1. The paper has a very large number of figures and subpanels. It could be useful to summarise some statistics into a table, to aid comparison between different models and metrics used for analysis. For example, providing tables for each region (or NES), that summarise the different statistics (e.g., ACC, MSSS, etc), at different lead times, for the different models (perhaps replacing figures 6-8)? This could help avoiding comparison between different figures and subpanels, and make the results clearer.
2. Do not use jet or rainbow for any figures. Use perceptually uniform colormaps, e.g., sticking to Red-Blue for all anomalies, and other options for sequential data (e.g., mean temperature).
3. Through the text, please clarify whether you use NES to refer to whole model domain, or just the shelf region. If you are more interested in model skill for former or latter, suggest clarifying this in both the introduction and summary.

Technical comments:

Section 2: Make sure you provide sufficient references to all CMEMS products – include reference dataset ID, DOI, and date of access. (Following recommended citation format). This is provided for some datasets, but not all.

L83: Clarify whether model levels, or post-processed product from CMEMS? (Or are these the same?)

L92: Why did you choose annual-mean rather than seasonal analysis? Please include further discussion on what impact this might have on results.

L92-93: The removal of climatology should be explained in more detail at the first point of introduction.

L106-109: Why the difference between use of products here?

L119-121: What about frequency of ocean lateral boundary forcing?

L126: Curvilinear rather than variable?

L128: For setting minimum depth, is this modifying the bathymetry or the free surface code?

L131: Remove first sentence (“realistic” is subjective) – just confirm choice of bathymetry product and its resolution.

L144: River data is from what data product – e.g., model or gauges?

L145: Why only M2 and S2?

L146: Clarify – ROMS-DOWN *forecast* simulations.

Section 2.4: Please clarify why different time periods used for each climatology. Also, why are monthly anomalies used rather than daily for ocean boundaries?

L154: Is the “mean” referring to monthly, or otherwise?

L159-161: Move this note on northern boundary location to support decision on domain size in previous section.

L173-175: By NES, clarify whether mean just shelf, or whole model domain.

L179-180: Are there no observations you could use for further comparison of surface or subsurface, e.g., SST, SSH, moorings, tide gauges? Please provide more justification and discussion on your choice.

L185: Clarify what LY1-LY8 refer to, e.g., for LY1, what start date used to predict 1997?

L219-221: It would be more beneficial to include the anomaly plots within the paper rather than mean climatology. Also, for SST and SSH, why only comparing with GLORYS, and not direct to obs? NB. GLORYS will use data assimilation – clarify whether same products used for comparison? Where are the anomaly maps for GLORYS vs Obs?

L230-231: Again, is the purpose of this model to reproduce shelf or off-shelf conditions? If shelf, should you focus on performance here, rather than whole domain?  

L236-237 & Figure 3: I don't understand what this figure shows - is it intergrated or mean RMSE?

L263: but does have a comparable or higher r for some lead times (as discussed later)?

L299-303: Could some of these summary scores be shown in a table instead, to help comparison between the different configurations and statistics?

L312: “Improved” reveals results, suggest delete or rephrase: Performance of ...

L328: Suggests improved, but comparing with GLORYS, which doesn't have tidal mixing?

Sections 4.1-4.2: These are results, suggest add separate results section to focus on regional applications/case studies?

L541: Rephrase: how basin-scale anomalies impact the NES. (*shelf* is within the acronym)

L545: This result appears quite stark and surprising that only introduced in detail here?

L556-557: More clarity of this would be useful in the methods section.

L558: Not sure I follow this reasoning, please clarify.  

L567-568: I agree it may not affect the broad conclusions. However, tides could alter the variability from mean state on annual timescales, especially at depth, if they affect the transfer of heat from surface and depth of mixed layers. This is worthy of more discussion in context of heat content and bottom temperature results.

L573: Need to work with partners to ensure required outputs are stored for downscaling – for both atmospheric forcing and ocean boundaries?

L583: Refer to variability rather than bias, to avoid confusion with performance stats?

L595: transferred rather than projected?

L620: Comment on how your skill scores <8y address this? Is the 2-5y timescale more useful to stakeholders?

Figure 1: Edit caption to clarify whether NES domain refers to white box or whole figure. Define MAB, GB and GOM, and also label clearly on figure.

Figure 2: Suggest showing anomalies rather than mean here, as easier to assess model performance. Also, please add bathymetry contours to show the location of the shelf break for reference. (This may be obvious for some variables, but not others. )

Figures 6-8: Suggest convert to table(s). If retain any, please label subpanels with titles to help reader. Also, why different colormap for ACC in Figure 8?

Figure 11: Please expand caption.

Figures 13 & 15: Again, consider conversion of ACC stats to table, to reduce number of figures.

Figure A6: Are all markers significant?