Dear Authors,

I read and reviewed the manuscript "Leveraging reforecasts for flood estimation with long continuous simulation: a proof-of-concept study". While the idea of the research is familiar (I have guided students on the same topic), and i understand the difficulties that play when applying such research in the Alpine domain. I find that the manuscript has too many messages which makes the end results to scattered. It would be wise to focus on one topic (e.g. meteorology) and leave out maybe some things to make the manuscript better stick (e.g. dynamic downscaling, hydrological modelling, etc. I leave this to the authors) . In the manuscript, one is often referred to the Supplement which is not a part of the manuscript. The manuscript could benefit from an additional experimental setup section where the different experiments that have been conducted to answer the research question are described. The research question is not clearly defined in my view.  The main issue with this type of work is if the data used is homogeneous and second if so, how the fusion of the trace is conducted and how this may affect results (effects of data assimilation in the first couple of days of the reforecast etc, leave out the first days of the RFs, etc). I didn't see that information in the manuscript.

Specifics:

• Did you check the data is homogeneous (likely it is not if you used all these different IFS cycles)? I didn’t see such an analysis
• Did you discard the first days or the forecast as they are affected by data assimilation? And if not did you check if this influenced the outcomes
• Mas et al 2023 is grey literature can you supply doi / link as it doesn’t seemto be available for others, supply a better reference which is available
• Figure 7 colour blind friendly? Difficult to see the dots maybe for some
• Too often referred to results in Supplement materials in my opinion
• Table 2 Is not mentioned in text
• Table 3 is not mentioned in text (and at strange position)???
• Leave out the paragraphs on dynamic downscaling read like a research report
• Figure 10 why compare against the world? distracts
• I would also expect return plots related to temperature/snow melt
• Why was TabsD not used for bias correction temperature? similar to RHiresD?
• Figure 8 make the observations more visible (black?)
• Conclusions are not very informative (what were the specific research questions?)
• Figure 9 colourblind friendly?
• Make data/code available in a repository
• Maybe focus on own results vs observations first before comparing against GWEX etc

The paper is very well written and provides insight into novel methods to estimate extreme values for floods.

I believe the manuscript overstates uncertainty reduction by implicitly treating the ~10 000-year RF dataset as a large set of independent samples. In reality, these are model-generated statistical draws from a numerical weather prediction system, sharing common physics, parameterizations, and structural errors. Increasing the number of RF realizations reduces Monte Carlo variability within that model, but does not reduce epistemic or structural uncertainty.

Uncertainty in the underlying weather model propagates directly into the flood estimates and this should be discussed. Systematic biases in precipitation extremes, storm persistence, and temperature–precipitation co-variability are inherited by the hydrological simulations and cannot be mitigated by ensemble size alone. Bias correction and stochastic downscaling adjust marginal statistics but do not correct errors in event dynamics, spatial coherence, or compound processes that control extreme floods.

As a result, the narrowing of exceedance curves reflects internal consistency of a fixed meteorological–hydrological modelling chain, not increased confidence in true flood return levels. The authors should explicitly state that RFs reduce sampling uncertainty conditional on one weather model and postprocessing framework, while leaving weather-model, structural, and climate-representation uncertainties largely unresolved.