This is an excellent study and important contribution to efforts making use of COS data to study the global carbon cycle. I have a few mostly minor comments:

The authors state the specific goal of exploring inversions using separate and combined datasets (L86-88), which I think is a great contribution. But I don't understand why bias free MIPAS only experiments were not included. It seems that if the combined inversion requires a bias correction to MIPAS, then the MIPAS-only inversion should also be based on corrected data. Please explain why or why not this was not included, and please include bias-corrected MIPAS-only inversion results if you think necessary. I also recommend to remove results based on uncorrected data from the abstract. You could simplify the abstract by mentioning the bias correction is necessary and performed, and limit the reported findings to those based on corrected data.

The final message in the abstract “better characterization of biosphere and ocean fluxes by observations is urgently needed” is mixed with respect to recommendations moving forward. There appear to be three messages: (1) better characterization of biosphere prior and uncertainty, (2) better characterization of combined ocean + land fluxes, and (3) better constraint of both by combining surface and satellite observations. It would help if the authors could state more clearly what is the conclusion based on the present study, and what set of steps need to be taken to bring more clarify to tropical COS fluxes. The finding in Section 4.4 (Line 418-419) that “inclusion of MIPAS data in the inversion leads to a better separation of land and ocean fluxes” seems like an important takeaway for the abstract.

Section 2.2: I think the choice of the 8^th level and high sensitivity AK (> 0.03) MIPAS data is fine. But this begs the question, how do the tradeoffs between data density and surface sensitivity affect inferred fluxes (spatial/seasonal distribution, land/ocean partitioning, etc)? The authors speculate about including higher level data in Section 5 (Line 479) but say nothing about lower levels. I suggest, if it’s not too much extra work and the justification is there, for the authors to conduct another run using lower level data, to explore changes in signal and uncertainty. Even at lower levels, the reduced amount of valid data will still overwhelm the NOAA data (e.g., 490 in 2009). At the very least, it’s worth discussing and speculating about these tradeoffs.

Line 33: Suggest referring to “atmosphere” somewhere in the description of the transport model (e.g., The transport model describes physical transport by atmospheric winds and chemical loss/production by atmospheric chemistry”)

Line 178: Are SIB4 fluxes year specific or climatology? Does this matter?

Line 201: Use of “generally the same” is vague. Are there important differences we should care about?

Section 5: I’m curious how much of the Amazon source can be attributed to outliers during a specific year of the inversion. The authors analyze the period 2008-2010, but I don’t see any mention of specific years. This could be a more robust finding if the inferred source is persistent over the three year period.  

The manuscript by Ma et al. constrained the global budget of COS using a combination of surface and satellite observations. They implemented a bias correction scheme to resolve the discrepancies between these two datasets and evaluated the results using independent aircraft and airborne data. Their results show consistent lower biosphere uptake of COS from the prior, especially in the Amazonia. The observation-constrained COS budget is an important topic and fits the ACP readership. The manuscript is well-written with nice experiment setups. I have minor comments that could potentially improve the clarity of the presentations.

My major concern is how robust are the strong reductions of COS sources over the Amazonia. The manuscript also questions the realism of this result. How sensitive are these results to the inversion parameters and observations? How does the result fit in the error bars? Is there any other work or independent data showing a similar conclusion. These additional tests and discussions are needed to better support the conclusion.

Specific comments

L11, is there any support for the claim that Amazonia is a COS source from other data? Otherwise, this would suggest that the posterior results are not realistic.

L15, what data are these errors evaluated against?

L55-72, it would be easier for readers to understand if the budget is shown in a table.

Eqn 6, how is “error” defined?

L231, why is a spin down period needed?

Eqn 7, how are the values in the satellite data compared with NOAA measurements? Are their differences consistent with the 0.003 error for beta here?

L253, how is this balance determined? Does it mean the two terms have similar magnitudes in the cost function? Please clarify.

Fig 3, why do you need a spin up and spin down for the constraint? How did you choose the length? It would be nice to add another line of MIPAS data at the NOAA locations.

Fig 6, it looks like the bias correction parameters are not adjusted to reduce biases in the tropical regions. Why is that?

Fig 9, the bars are too thin to be easily identified. Please adjust the width.

491, are these related to the specification of errors? Would the results improve if changing prior error characterizations?