This manuscript investigates the 2019 Brazilian fire season under non-extreme climate conditions using a two-step CTE-LW/SW inversion framework constrained by TROPOMI and MOPITT XCO observations. The topic is scientifically relevant and timely. The inversion design is carefully structured, and several sensitivity experiments are conducted to assess robustness. The study provides valuable insights into regional fire CO emissions and bottom-up inventory discrepancies.

The manuscript is generally well organized and supported by data. However, several methodological aspects require further clarification and quantitative support before publication.

Major Comments

1. Identifiability of the Two-Step Inversion Framework

The manuscript applies LW to optimize background CO and SW to optimize fire emissions. However: How does the LW step avoid absorbing part of the fire signal?

Does LW optimization alter OH or NMVOC production patterns regionally?

Could the SW step compensate for residual background bias?

The manuscript mentions a remaining ~20 ppb positive bias in the free troposphere but argues it does not dominate SW results. This claim requires quantitative sensitivity tests. Please clarify information flow and provide supporting diagnostics.

2. Justification of the XCO ≥125 ppb Filtering Threshold

The chosen threshold is used to isolate fire plumes, but: How sensitive are total emissions to this threshold? Could low-intensity or diffuse fires be excluded? Please provide emission estimates under different thresholds (e.g., 100, 125, 150 ppb). Also include statistics of excluded observations (<125 ppb) to demonstrate that the filtering does not bias total fire emissions.

3. Observation Error Inflation Strategy

TROPOMI errors are inflated (transport error doubled + 4 ppb added) to achieve χ² ≈ 1. While common in data assimilation, further clarification is needed:Show χ² distributions before and after inflation (not only the mean). Explain how the “×2 + 4 ppb” parameters were chosen. Provide sensitivity tests to alternative inflation factors. Importantly, TROPOMI native uncertainty (~3 ppb) is much smaller than MOPITT (~7 ppb). Inflation changes their relative weighting. Please discuss whether the reported convergence between TROPOMI and MOPITT inversions is partly influenced by this weighting adjustment.

4. CO-to-C Conversion Assumption

CO emissions are converted to carbon emissions using a fixed CO/CO₂ ratio. However, emission factors vary by fuel type and biome. Please provide sensitivity analysis (e.g., ±20% in CO/CO₂ ratio). Quantify the resulting uncertainty in carbon estimates. Clarify the limitations of using CO-only inversion for total carbon quantification.

5. Underestimation over Cerrado/Caatinga

The manuscript reports posterior emissions nearly twice the inventory values in this region. While fuel pool explanations are discussed, the attribution remains qualitative.

Please consider: Showing spatial posterior/prior ratios. Providing innovation statistics by biome. Clarifying whether transport redistribution contributes to the regional signal.

Minor Comments

1.Clarify how the two-month MOPITT data gap in 2019 affects results.

2.State earlier in the methods that the system cannot constrain super-local plumes due to resolution limits.

3.Clarify the origin of NMVOC and CH₄ pre-scaling factors.

4.In the conclusions, better define the applicable scale (regional totals vs. event-scale emissions) to avoid overgeneralization.

Technical / Presentation Issues

1.The “thick semi-transparent black line” in Figure 1 is difficult to identify; please adjust visibility.

2.Provide bias and RMSE metrics for ATTO validation.

3.Add y-axis labels in Figure 5 for clarity.

Constraining fire carbon emissions is essential for understanding the carbon budget and carbon cycle. This study presents satellite-based estimates of CO emissions from fires in Brazil using inverse modeling. Different inversion configurations lead to consistent posterior estimates, demonstrating the robustness of the framework. The manuscript is generally well-structured and clearly written. This study demonstrates physical consistency of fire emission monitoring through an innovative two-step inversion framework (long-window and short-window inversion), multi-source satellite data fusion, dynamic error modeling, and detailed biome-level analysis. It not only provides reliable emission estimates for the 2019 fires in Brazil but also offers important methodological references for the design of future fire monitoring systems, the improvement of emission inventories, and carbon cycle research. The manuscript is valuable and suitable for publication after moderate revisions. My comments are as follows:

Comments

1. The entireframework hinges on the premise that the long-window (LW) inversion can effectively optimize non-fire sources (NMVOC oxidation, CH4 oxidation, anthropogenic) to provide a pristine background for the short-window (SW) inversion, which then exclusively optimizes fire emissions. There is a risk of "aliasing," where fire signals not fully captured by the prior fire inventory could be absorbed by the LW-optimized terms (e.g., through adjustments to NMVOC production), or conversely, that residual errors in the LW background could be erroneously attributed to fires in the SW step. The authors should include a dedicated sensitivity analysis or a set of diagnostics to validate this decoupling. 
2. What is the practical benefit of the LW inversion in this application? While correcting background concentrations is conceptually beneficial, Fig. 2b shows substantial background bias distributed across the entire domain rather than localized in fire regions, which may limit the effectiveness of the LW inversion. Please further illustrate the prior residuals before applying the LW inversion to clarify its contribution.
3. MOPITT data are missing between 26 June and 24 August (L203), covering roughly one quarter of the inversion period. Since Fig. 5 shows that deforestation fire emissions peak in mid-August, this gap may introduce non-negligible bias. Notably, the posterior mean is averaged across three inversions, whereas GFED5.1_MOPITT represents the prior before 24 August only. This could shift the posterior mean of deforestation fires toward the GFED5.1 prior (see Fig. A4). Please clarify this issue.
4. 2.5.1: How are observations within the 3-day assimilation window treated, given that data are available at roughly three timesteps?
5. Line 248: An outlier threshold of three is standard in many inversion systems. However, for rapidly varying fire emissions, large model-data mismatches at individual timesteps may result in the rejection of valid observations. The missing peaks in Fig. 3 suggest this may be occurring. Please discuss the potential impact of the filtering strategy.
6. 2: It would be helpful to include the spatial pattern of prior emissions to facilitate comparison between emission and concentration patterns.
7. Lines 404–405: Fig. 2d does not present an observation-simulation comparison. Fig. 2e shows residuals, but residuals over B1 remain substantial. Please clarify or correct the description.
8. The inversion domain extends beyond Brazil (Fig. 1), yet only Brazilian results are discussed. Given the significant CO fire signals in surrounding regions (Fig. 2a), how do emission adjustments in neighboring countries influence the Brazilian estimates?
9. Relatedly, could biases in emissions from surrounding biomes contribute to the inferred increases in the Cerrado and Caatinga? For example, Fig. 2b shows substantial underestimation over deforestation and Pantanal regions, which would likely drive emission increases during inversion. Although a biome-dependent covariance structure may limit such cross-region influence, the Amazon savanna fire region is not separated in Fig. 1.
10. The authors do a good job explaining the inflation of the model-data mismatch to achieve a χ² near 1. It would be useful to state the mean χ² value afterinflation, not just that it "approached one." Was it 1.05? 0.98? Providing the number adds precision.
11. Author presents a clear mass balance equation. I may miss it somewhere, the cost of function and configuration of the prior error covariance matrix B is not clearly described. Given that the specification of B can significantly influence the posterior results, the authors should provide a detailed explanation of its construction (e.g., variances and auto-correlations).