Gerneral commnets

The manuscript presents "RTSEvo v1.0," a dynamic evolution model for Retrogressive Thaw Slumps (RTS). The authors address a critical gap in current research by advancing from static susceptibility mapping to spatiotemporal simulation. I find the methodological framework to be highly innovative. This hybrid approach offers valuable insights into the morphological evolution of RTS and significantly contributes to our understanding of abrupt thaw processes and permafrost degradation mechanisms on the Qinghai-Tibet Plateau. The study is well-structured and tackles a timely issue in permafrost science. However, I have several concerns regarding the spatial scale of the input data, the transferability of the model, and the standardization of the source code that need to be addressed before publication.

Major comments

1. Scale Mismatch and Resolution Limitations: Retrogressive Thaw Slumps (RTS) are local-scale periglacial landforms. In this study, the selected RTS features have an average area of 2.61 ha, which corresponds to an approximate diameter of 160 m (assuming a square geometry). However, the spatial resolution of several predictor variables used in the model is relatively coarse; for instance, the NDVI data is at 250 m resolution. Even utilizing 30 m resolution datasets may be insufficient for simulating such small-scale geomorphological processes. Consequently, this "scale gap" represents a significant constraint on the model's precision. I strongly recommend that the authors include a serious discussion regarding this limitation in the manuscript.

2: Model Transferability and Generalizability The study constructs the RTSEvo model by combining physical process characterization with machine learning parameter calibration. While the model was validated using over 450 RTS inventory records with positive results, the spatial distribution of these samples is highly concentrated within the Beiluhe Basin. This spatial clustering implies that the climatic conditions, which are key drivers used to predict RTS state and evolution，are highly homogeneous across the training and validation datasets. Therefore, it is questionable whether the parameters calibrated in this specific region are transferable to other permafrost regions with different environmental settings. The authors must strictly clarify the model's transferability and discuss this potential lack of generalizability in the text.

3 Codes: Geoscientific Model Development (GMD) mandates that code be open-source and sufficiently documented to facilitate reuse by the community (and for the purpose of peer review). Upon reviewing the provided repository. I found that the current documentation is non-standard and lacks rigor. Notably, the code contains comments in languages other than English (e.g., Chinese characters are visible in comments. I strongly suggest the authors perform a thorough revision of the code and documentation to ensure it meets international standards and undergoes rigorous testing before final publication.

Specific comments

Figure A2: Regarding the final plot in Figure A2 (Active Layer Thickness): Why does the distribution exhibit a bimodal pattern with values clustered at two extremes (showing a difference of up to 1 m)? Please clarify the physical or data-processing reason for this distinct separation.

Xu et al.’s work fills a critical bottleneck in permafrost research by moving from static susceptibility assessment of RTS to dynamic modeling which can be used to predict the future development of RTS. This represents an outstanding contribution to the field. Overall, the proposed framework is well constructed, the modeling strategy is clearly implemented, and the results show reasonable skill in reproducing the observed patterns of RTS development. However, I also have some concerns regarding scale, sampling, and long-term validity require further clarification.

Major comments

1. The model operates at a 10 m spatial resolution, while many key driving variables are resampled from much coarse resolutions (e.g., 1km). Please discuss how this resampling affects the precision of the probability estimation and the subsequent spatial allocation module.
2. The FoM values appear numerically not very high. I know this framework adapted from LUCC modeling, please provide a comparison indicating if these metrics are consistent with or superior to those reported in related LUCC or other geohazard evolution studies.
3. The model assumes that RTS state transitions are irreversible (line 459), which is noted as a limitation for long-term projections. Please clarify the physical justification for maintaining the assumption in v1.0 and discuss if a recovery state could be integrated in future versions. Furthermore, what is the specific temporal horizon (e.g., 10 vs. 50 years) over which the current model remains a reliable predictive tool?
4. The use of 1:1 sampling ratio for RTS vs. non-RTS pixels (line 234) doesn't reflect the natural rarity of RTS in the landscape. Please address how this balanced training approach affects the calibration of the final probability maps and if any threshold adjustments were required.

Specific Comments

1. Line 62: “guid the simulation” should be “guide the simulation”.
2. Line 461: "longer temporal horizons" could be more specific (e.g., decadal to centennial scales).
3. Line 433: Provide the full name of "SHAP" upon its first occurrence.