Summary

The study by Mahmood et al. makes an important step forward in understanding initialized prediction on decadal timescales (out to 30 years), a topic that is only rarely addressed due to computation costs. The results are very interesting, and ultimately seem to hinge on the fact that the initialized runs push AMOC into a weakened state that is outside the range of what is captured by the uninitialized simulations. The authors could enhance the impact of this study by focusing in a bit more on this topic – is this a realistic plausible reality? Why do these differences exist in the Lab Sea convection? And if initialized predictions are perhaps unreliable at this timescale (multi-decadal), should the community be pushing instead towards what the authors refer to as “variability-constrained projections” (which would benefit from even a brief description).

Major Comments:

• The discussion in the first paragraph of Section 3.1 should much more nuanced. The authors state that most land areas show ACC that is positive and statistically significant, with an exception of South America; But looking at Fig 1, there is a lot of stippling indicating things are not significant, even over land. Signals over Australia, for example, are rarely significant. In comparing with Fig S1, again it seems that significance has actually dropped dramatically in the case with fewer years (most notable over the Indian Ocean and Pacific). This calls into question the assertion that “a reduced sample size (due to initializing by every fifth year) does not strongly affect the skill of the prediction system.”
• Figure 1: Given how much of the signal is not significant and thus stippled, I wonder if the fig would be more readable if stippling indicated where things are significant? It’s hard to discern at first glance right now.
• Discussion of Fig 2 notes the different mean state in initialized predictions, but should also discuss the much larger bias as a result.
• The discussion around the Labrador Sea is interesting, as it points to inherent differences between initialized and uninitialized models. Some added context would strengthen the section further; for instance, is a shutdown in the Lab Sea convection rooted in any observations? Is it realistic, or an unwanted model behavior? Why does this happen in initialized runs?
• Line 305: “The climate predictions based on these variability-constrained projections (e.g. Mahmood et al., 2022; Donat et al., 2024) even show higher and more widespread added skill, compared to the initialized predictions presented in this study, for both 10 and 20 year mean predictions.” – this seems like a huge point, but isn’t backed up by comparison within this text. It would make a stronger paper overall to focus in on this as a piece of the larger manuscript.

Minor Comments

• Line 51: What is meant by “constraining patterns of climate variability in large ensembles of the uninitialized projection simulations”? Some added description would help.
• Line 79: “…initialized on the first of November every 5^th year…” – is there any rationale for using November? Is this consistent with other experiments? Would differences be expected if initializing in spring/summer instead of fall/winter? (Perhaps the Lab Sea response would differ in particular?)
• Line 90: “…on the order of 10-5 K)” – to clarify, is this meant to be 1e-5? That seems like it might actually be rather large, but I’m typically working in uninitialized space, wherein 1e-14 perturbations are common. Could the authors comment on if this is a standard magnitude for initialized predictions?
• Line 106: “…to a uniform five-degree grid” – that’s really coarse; is there a reason to convert to five degrees? Is this the native resolution of some runs vs others?
• Line 141: “…evaluation periods are different (i.e. 1961-2020 for FY1-10, 1971-2020 for FY11-20 and 1971-2020 for FY21-29).” – Could you limit the evaluation period for FY1-10 to 1971 onwards arbitrarily, just to test the impact?
• Line 149: “…over land regions including northwest Canada and USA.” – I’m not sure I see a lot of added value over the USA, it’s mostly stippled? But Australia and the Middle East seem to show something more cohesive.
• Line 162: “Previous studies have shown that the decadal predictions can be highly skillful in forecasting temperature over SPNA region.” – Citations?
• Fig 2: Please add a legend for the lines, in addition to the description the caption for easier readability.
• Fig 5: Suggest adding stippling for significance, as in other maps

The added value of initialization for long (multidecadal) climate outlooks remains unclear, so this analysis of 30-year initialized predictions using the EC-Earth3 model will be of considerable interest to the decadal prediction community. The manuscript is clear and concise, and the results are intriguing.  I recommend this be published after some minor issues are addressed. Most importantly, I think an additional analysis is needed (perhaps in supplemental) that shows how skill varies with lead time when a common verification window is used.

Minor Comments

L48-60: A recent study suggests that initialized internal variability could constrain surface climate variability for multiple decades (Deser et al., 2025, https://doi.org/10.1007/s00382-024-07553-z). It may be worth including such potential predictability examples in this intro section.

L62: Unclear what “model dependence in previous results” refers to exactly

L112: If hindcasts are 30 years long, why not examine FY21-30 and FY11-30, for consistency?

L121: Should mention that observed anomalies are computed in similar fashion (observed climatology matching months/years used for hindcast climatology).

Fig. 1: Caption should state the field being analyzed. Methods should clarify how statistical significance is quantified.

L137: Authors have an opportunity here to directly address skill sensitivities to sampling and system design choices, so why not do that in Fig. S1? A first FY1-10 ACC plot could subsample the previous system, selecting only start dates used in Fig. 1 (this would isolate the effects of changes in initialization, I believe). A second FY1-10 ACC plot could show skill when all start dates are included from the previous system (which would reveal the effect of reduced hindcast sampling on skill).

L142: I think this should read “1981-2020 for FY21-29”? As noted above, the use of 9-year average (instead of 10-year average) for FY21-29 introduces another avoidable difference. The differences in verification window are likely important, so why not include a supplemental figure that shows skill for a common verification window (1981-2020 for all leads)? Otherwise, the conclusions that can be drawn from this comparison are very limited.

L152: I think the conclusions about impact of initialization are somewhat shaky given the discrepancies in verification window (see above).

L187: I would delete “as it was assumed in the past”, because I don’t consider fast convergence to be a generally held assumption, particularly for the North Atlantic.

L213-215: Repetitive with introduction.

Fig. 4: Is it surprising that the AMOC climatology in DP_30yrs is almost identical to that from the DP (10-year), even at short leads, despite the different initialization methods?