This manuscript addresses a well-recognized issue in ground-based solar radiation QC: The globally uniform thresholds of traditional ERL tests are overly conservative and frequently insensitive to localized environmental variations. Through a newly proposed method, this work effectively tightens conservative limits. This paper is well-written, logically sound, and worthy of publication in AMT after minor revisions. Below are my detailed comments.

1. Eqs (1)-(3) and Eqs. (5)-(7) both contain the cosine of zenith angle. Would it be clearer if the notations are standardized? The same applies to the notation in Line 352.
2. In Section 2.2.1 and Fig. 2 of the manuscript, there is an inconsistency regarding the feature variable abbreviations. While the figure caption denotes the feature variable as B_n, the corresponding label inside the actual illustration of Fig. 2 is marked as B_h.
3. It is unclear to me why the number of harmonics is set to 25. If it is simply the yearly cycles that need to be identified, a few harmonics should work, given the regularity of the seasonal component.
4. In Section 3.2, the authors utilize the L-BFGS-B algorithm to optimize the coefficients of the structural limit curves. However, the choice of the initial guess for this iterative solver is not specified.
5. The objective function for parameter optimization is designed to maximize the F1-score. While this is mathematically reasonable, the manuscript would benefit from a brief discussion on its meteorological and practical implications. Specifically, the authors should explicitly justify why a balanced F1-score is preferred in this radiation QC framework.
6. From Table 3, it shows that the values of parameter c_G are often very small, i.e., near zero. This is inconsistent with the original test, where the parameter c should reflect the measurement uncertainty, as also mentioned by the authors in the introduction. At the same time, the values of c_Bare capped at 10. This suggests an issue with the optimization setup.
7. The authors should explain how the proposed radiation classification can be applied to other locations.
8. The authors are expected to append some qualitative scientific remarks in the discussion or future outlook section to enhance the comprehensiveness of this methodological framework. For instance, can this approach be used to handle other QC tests, such as the closure test?
9. Comparing Fig. 7 (Class 6, B_n) with Fig. A6 (top two panels), the scatter plots look different. Why? It seems to me that the latter contain more points.