The study of Smallmann et al. presents a model-data integration study where a suite of terrestrial ecosystem models of increasing complexity is inverted and evaluated using spatially resolved data across Brazil. The finding that already with quite simple models, parameter uncertainty is more important than model structural uncertainty and uncertainty in forcing data has a large impact of the earth system science and is worth to be published.

I enjoyed reading the beginning of the paper and appreciated the well designed study using multi-model, multi-biome, site-specific spatially resolved setup, and varying input data for future climate scenarios, within a fully Bayesian inversion setting.

However, I got disappointed when I more closely inspected Table 2. Even with the simplest model, the confidence intervals of the predictions are so large, that only vague and general statements or conclusions can be drawn from the results. All the elaborations and conclusions on model complexity and structural error would be a really good presentations, if the results were more constrained. However, with the large uncertainties I would recommend to only summarize them and omit the detailed presentation, because they are base on vague ground. Instead of studying increasing model complexity, the results show that the data is already not enough to constrain the simplest model variant. Statements like LL259 “simulated NEE was consistent with CTE ensemble at the 90% CI” does not tell me much about the goodness of the model, if the CI range is 400% of the median estimate. In order to defend the insights despite the large uncertainty, the posterior density of the parameters in comparison to the priors should be provided as a supplement or appendix.

The main conclusion about the dominance of parameter uncertainty is strengthened by this large uncertainty and should be published, with a much shortened presentation of the (to my opinion vague) comparison across model structures.

An alternative route, which requires a larger reanalysis effort, is based on the claim of the authors that the model can be constrained by repeated EO observations of biomass. In addition to the current model inversion, I suggest generating an artificial observation of this biomass data stream using the most complex model variant add noise and some slowly-changing bias and repeat the inversion including this artificial data. If the uncertainties decrease as much, the presentation about model structure could be kept, but based on this new (artificially) more  constrained inversion results.

Specific comments

To gain an conception about the computational effort: At how many pixels was the model inverted?

Line 220ff: It did understand how “future climate is imposed by determining the anomaly from the end of the analysis until 2100”. Please, extend this explanation.

Line 223: I assume the model structural uncertainty was estimated for each climate scenario separately (and the climate uncertainty for each model variant separately), right? Or does the “between model range” span across all climate scenarios?

Fig 4: In my opinion, the stippling (indicating consistency within confidence range) does not tell much when considering the large uncertainties.

Fig 5: putting all the labels the center panel confused me first, I suggest putting the observation legends to the panels. Almost all the streams are encoded by color, which for me were difficult to read.

L286, 294: Why is the NEE not improved with model complexity, if fire is improved and makes up 3 to 30% of NEE?

L315: How did you assign priors to the MRT parameter? The sentence suggests a Normal distribution that includes also negative values. A lognormal prior would be more appropriate and you could report the multiplicative moments of the posterior and avoid negative residence times.

Sec 3.3. reads lengthy. Are all the details necessary in the main text. I have, though, no specific suggestion how to shorten.

L331: Hints to model error. Thanks for the discussion at L462ff, that could be referenced at this point. For me it did not become clear, how biomass removal was accounted for in the DALEC simulations, and the future scenarios.

L 347: Can you quantify “most likely”? Can you infer p(deltaBiomass(t) > 0) from the posterior?

L 450: may replace “a function of three factors” by “There are three possible interacting explanations”

The manuscript "Parameter uncertainty dominates C cycle forecast errors over most of Brazil for the 21st Century" by Smallman et al. presents projections of Brazilian carbon cycling using a model-data hybrid approach. A terrestrial ecosystem model with varying degrees of complexity is constrained by remote sensing data before being used to quantify the evolution of carbon stocks into the future. Overall, I think the methodology is robust and I commend the approach, in particular separating and quantifying parameter, structural, and climate uncertainty on future stocks. The paper is well written, flows well, and addresses an important topic with substantial and robust conclusions. 

Therefore, I would recommend the manuscript to be published after the following comments have been addressed:

1. In general, the paper has detailed reporting of results, however the explanations of model behaviour are sometimes incomplete. This lets down the previous good work that comes beforehand. For example:

   L406 - Can you explain why RT increased?

   L407 - I find it interesting that here the net flux correlates with GPP, but the long term net changes have low correlation with GPP. Can you explain?

   L414 - Can you explain why Amazon flux is overestimated?

   It would help the reader if more insight was offered into what drives model responses.

2. The introduction stresses the importance of land-use change and fire for the Brazilian carbon cycle, but there is little discussion on how these disturbances will influence future carbon stocks. Are they important? If not, why not? If they are, then how important? Also, it is not clear if/how land-use change is implemented in your future simulations. Is this implicit in the mortality parameter? It would be good to clarify this.

3. The writing style can be improved in places. There are a few dense paragraphs (that can sometimes read like a list), e.g. in Sections 3.3 and 3.4.3, there is often a detailed comparison between the five models that doesn't add much value. Can the values be moved to a table and then the text just discusses the most important parts? E.g. "Carbon allocation to wood as well as wood turnover rates determine future biomass stocks... Relative contribution of allocation and turnover varies across biomes…". Also, at times the paper sounds quite negative. E.g. in Section 3.4, I feel some value is lost when all your results are caveated with "uncertainty is larger than predicted changes" or (L299-305) - "...insufficient observational constraint to confidently determine the sign of NBE or soil C dynamics. The same was largely true for wood stock dynamics..". I appreciate the uncertainty is important (and the focus of the paper), and certainly needs to be addressed, but this large uncertainty seems to undermine any projections you make about future stocks.

4. Can M1-M5 capture the full extent of model structural uncertainty? You start to discuss this (L467) but in my opinion, I don't think this is a strong enough justification. Further, you state "...limited sensitivity of C cycling to future soil moisture stress.", which is at odds with drought experiments that suggest increased mortality and reduced productivity (e.g. ACL da Costa et al., 2010 - New Phytologist).

Minor comments:

 - Looking at Figure A7, M5 looks like it still has a large fire bias similar to the other models, so can 

Can M5 say anything about the role of fire on future C stocks in the future?

 - FLUXCOM-RS+METEO is known to underestimate IAV. One option would be to use FLUXCOM-RS which uses interannual satellite LAI as an input and generally shows large IAV and trends. This could improve the comparison in Figure 5.

L46-47 - Missing words?

L299 - "Carbon dynamics of wood, not soil, is the primary driver of net exchange." - Does this depend on timescales? Over the course of a century, won't slower (but potentially larger) soil C dynamics play a bigger role?

L394 - Missing words?

L395 - Increased from what? Due to increasing model complexity? Over time?

L394-402 - It would be good to place wood litter in the context of other drivers of fire. Is wood litter the most important variable if we want to correctly simulate fire? How about the local climate? Ignition sources?

L407 - Do you mean "GPP" instead of "net carbon uptake"? 

L454 - Why is it important that NPPWood is the second most important determinant?

L456 - Don't some transient biomass maps exist already? e.g. ESA, or VOD based (Liu et al. 2015). 

L533 - But don't we have transient AGB and atmospheric NEE available now?