In this article, Kikaj et al. present in great detail the workflow to operate ANSTO dual-flow-loop two-filter radon detectors, and to process data to achieve best-quality radon observation suitable for atmospheric transport model validation and greenhouse gas (GHG) emission estimation (among many other goals, not necessarily mentionned by the authors). The authors finally provide recommendations that could be used to establish standard operation procedures in international cooperation programmes such as ICOS or GAW.

General appreciation

This article will be of undeniable value for the growing scientific community operating two-filter radon detectors across the world, and for the dissemination of high-quality radon data for various scientific or operational uses. The article is fully in the scope of Atmospheric Measurement Techniques and meets the required quality standards of the journal, both in terms of scientific robustness and presentation quality.

Therefore I recommend the publication of this manuscript. Before this, minor revisions could be considered. Especially, some aspects pointed in my comments make the article sometimes hard to understand, and deserve to be clarified. Numerical values of instrument- or site-specific parameters are also sometimes missing.

Specific comments

Line 20 : « to keep this ... »→ « to keep the ... » ?

Lines 20-55 : This is a long (but nevertheless interesting !) introduction to justify the need of high-quality radon measurements. But this need would still be evident even introduced more briefly. So, was the manuscript to be shortened, could these paragraphs be reduced. Beyond the validation of ATMs and the estimation of GHG emissions, radon data have many other interesting application fields in atmospheric sciences. I miss a bit of a wider view on applications while reading this introduction (and the final conclusions).

Line 141 : Does « local time » refer to the official time in the UK (including winter/summer changes), or the local solar time (i.e. UTC)? The latter is much more relevant when considering diurnal cycles in the planetary boundary layer. Please specify.

Line 163 : Is WAO a global or regional GAW site?

Line 232 : What is the empirical procedure to determine an appropriate flushing time ? What are the criteria ? Please go into more detail.

Figure 2 : It seems that this figure uses data from a real calibration. Nevertheless the linear trend estimated from outdoor ambient air before and after the calibration is hardly visible. For clarity, could a schematic be presented instead of real data ? Further, I suggest that the variable names LLD_peak and LLD_peak,a should be displayed in the figure or in the legend. Finally, the figure should be enlarged (legend characters too small).

Lines 251-252, « This scaling factor will vary with the length of the calibration injection and sampling flow rate. » : What are the values retained for the 3 stations ? More generally, what is the prodedure to determine the scaling factor ?

Line 243 : The full-text development of the acronyms LLD and ULD is too discreetly hidden in Table A2. Please define these important variables in the main text body.

Lines 279-280 : It is not clear why a high calibration accuracy is needed to derive radon vertical gradients from tall towers. Could you provide values to support this statement (e.g. typical radon vertical gradients over 100 m vs. typical measurement accuracy)  ?

Lines 301-302, « the instrument undergoes few measurement cycles ... » : What is here meant as a « cycle » ? A 30-min count sequence ? I am afraid this term could be a bit confusing here, please rephrase.

Lines 318-319 : The inlet heights could be provided in Sections 2.1.1 and 2.1.2.

Lines 344-345, « Totalized counts include (…) and sample flow rate » : this sounds strange to me. LLD and ULD are clearly totalized counts, but how can it be so for a flow rate ? Do you mean that the measured flow rate is integrated over 30 min and recorded as a total air volume? Please clarify.

Figure 3 : There is a legend concerning values in red, green and yellow, but nothing is said about the parameters in bold black.

Lines 361-362, « a flow rate of 5.5 m s^-1 » : Usually flow rates are expressed as volumes (or masses) per unit time. Why is it expressed as a velocity here ? What is the section area needed to convert this air velocity into m³ s^-1 ?

Line 376 : What is a UPS ? Please write this acronym in full text.

Line 396, « if the blowers restarted early » : do you mean that the blowers may unexpectedly restart earlier than planned ? This sentence sounds strange to me.

Lines 396-397, « a check was made that the remaining data were linear (approximately constant) with relatively low variance » : Background counts are expected to show no trend within the retained measurement sequence, but to remain constant (only subject to count noise). The adjective « linear » is here confusing. Further, the count variance should be low relative to what ? What is the retained value for a variance threshold ?

Lines 407-408, « a site-specific threshold value » : Is the threshold value of 50 counts min^-1 valid for the three sites ? It would be useful if the authors could provide a method to determine such a threshhold at any site.

Section 3.3 : I must confess I had a hard time while reading this subsection, and am still not able to reproduce the fitting algorithm based on the details give here. With a concern of reproductibility, I think that the algorithm should explained in more detail and clarity, possibly as supplemental material. Alternatively, a well-documented routine (in python or other freeware language) could be made available on a public repository. More specifically :

• Lines 421 and ff. : Could a literature reference be provided on the Savitzky-Golay method ? As far as I know, this smoothing method consists of fitting a n^th-degree plynomial on the data contained in a moving time window centred around the current timestep. The result is a single value equal to the value of the polynomial at this timestamp (with n=0, the method is thus a simple moving averaging). The whole process is done again considering a new time window centered around the next timestamp, and so forth. For this reason, the number of data points in the resulting smoothed signal is equal to the number of raw data points. But in the present study an interpolated curve is obviously produced between the timestamps of the raw data points, which requires to handle several polynomials produced within adjacent time windows. It seems that information elements on this aspect are given in Lines 432-434, but I do not understand those elements, sorry. What is done in steps 3-5 of the procedure is also unclear to me. All this should be clarified, or at least appropriate references should be given.
• Lines 423-424 : After some effort I eventually understood that a 184-day (~ 6 months) moving window centred on the current timestamp is used. If I am right, couldn’t you express this more simply than « two half-regression windows, each with a duration of 92 days » ?
• Line 427 : Why does the result of the moving fit requires further smoothing ? This does make no sense to me since the result of the moving fit is a linear curve (1st order polynomial). Please clarify.
• Line 431 : Does « supplemented » should be understood as « extrapolated » ?
• Lines 442-442 : This is sophisticated way to express more simply that missing points are linearly interpolated between the closest available data points, and spaced evenly every gap_max days. Did I miss something more subtle ? Beyond this, what is the value of gap_max in your study ?
• Lines 444-445 and ff. : Annual reductions in sensitivity are expressed as percentages. How are these reduction percentages defined from the fitted calibration coefficients ? As the reduction of the value relative to the preceeding year ? Please specify.

Section 3.4 : The signal deconvolution to correct the slow instrument response (compared e.g. to GHG analyzers) is major aspect of the data processing. At first glance the algorithm is described in some detail in Griffiths et al. 2016, but I found no mention in this article, neither in the present manuscript, of code availability. Would routines be accessible somewhere ? This would be a considerable gain of time and effort for users.

Line 468, « the median of the deconvolution result at that timestamp » : this suggests the deconvolution result at a given timestamp is not a single radon volumic activity value. Further in the text (line 509), it is indeed written that the deconvolution process results in a statistical distribution. This is an important - but not straightforward! - point that should be clarified as early as here in the text.

Lines 501-502, « The maximum discrepancy (…) of 7.7%. » : This sentence is unclear. Should one understand that the discrepancy is of 7.7%? And 7.7% of what relative to what ? Please clarify.

Section 3.7 : The whole subsection deserves to be clarified. It is entitled « Combined measurement uncertainty », and lists and quantifies (as induced relative uncertainties) different specific source of uncertainty. But in the end, no numerical value is given for the resulting combined uncertainty. It is nevertheless stated in the very first sentence that the [combined] uncertainty is eventually derived from the distribution resulting from the deconvolution process (as the difference between the 16th and 84th percentiles). That the different uncertainty sources listed in this section affect this distribution is far from being evident to most readers, I am afraid. Could it be explained why? Are the listed relative uncertainties the specific impact of each source on this final percentile difference ? How are these values estimated ? It could also be clarified that the combined uncertainty is here not obtained as a rooted sum of squares of individual uncertainties, as more usually done.

Line 511, spelling : « Poisson ». Please specify how an uncertainty value can be derived from the Poisson distribution.

Lines 517-518 (« plate-out of unattached radon progeny ») and 522 (« plate-out effect ») : The term « plate-out » is obscure to me (it seems it has to do with particle deposition), and I don’t know exactly what the plate-out effect is. Could a short explanation and/or a reference be given?

Line 539, typo : 8^th → 84^th

Section 4 : This section is very convincing about the capability of the deconvolution process to account for sharp changes in ambient radon with the right timing (by comparing the radon signal to GES concentrations obtained with fast analyzers), as well as for absolute values of radon volumic activity in controlled conditions. However, it could be thought that the deconvolution algorithm had already undergone careful validation in Griffiths et al. 2016. Could you specify what is new concerning the validation checks made in the present study compared to the original paper ?

Lines 557-558 : This sentence would be clearer if « (²²²Rn_ini) » was moved just after « radon concentrations ».

Figure 6 : The panels in this figure are too small. They could be enlarged in order to fill the full text width. A thin vertical dashed line could be added at 9:00 LT, when the signal reaches 85 % of the target value. It also seems there is confusion in colors between the figure and the text (lines 560-562). Especially, the deconvolved signal is not purple (cf. line 562) but obviously yellow, and vice versa. Also, the target square waves could be displayed in the graphs for direct comparison with the deconvolved signal.

Line 594 : Change to read « (see Figure 10 in Griffiths et al. 2016) ». If the manuscript has been prepared with Latex, please consider using \citep[][]{} to avoid double parentheses (here and at some other places in the text), e.g. in this case : \citep[see Figure 10 in ][]{Griffiths2016}.

Line 601 : probably missing dash after « summer ».

Lines 619-620 : If during the night the measurements at those sites are decoupled from the surface (the inlets being above the inversion), how the CO₂ exhalated by the soil and accumulated near the surface could reach the inlets 3h before the radon, which also accumulated in similar conditions below the inversion? I am not conviced by the explanation given here for this 3h delay. Should an additional source of CO₂ be hypothesized?

Figure 8 : The interquartile range for radon is displayed in light yellow and is hardly visible. Despite the use of  color transparency, the superposition of variability strips is confusing. The variabilty of radon could be alternatively displayed e.g. with thin dashed curves.

Line 628-631 : For this coastal site nothing is said about the possible influence of land/see breezes. Could breezes play a role in the radon and GHG diurnal cycles ?

Figure 9 : the CO₂ curve at 50 m agl could be integrated to the central panel in Figure 8, and in turn Figure 9 could be removed.

Section 5 : The recommendation summarized in this section will be very helpful for operators of ANSTO radon detectors, thank you for them!

Line 687 : « three stages » is not useful in this subsection title.

Lines 710-712 : It is not clear what is the difference, in term of performance, between the Rust- and Python-based codes. It sounds like if the Python-code is less efficient or gives poorer results. Is it the case ? Could you clarify ?

Line 712, typo : « … the the … »

Line 739 : That this measurement protocol could be a significant contribution to climate change mitigation and the achievement of the Paris Agreement is a far reaching conclusion! This nice and very useful article would not suffer to have its final point after « processes ».

I again thank the authors for their work.

Summary:

Kikaj et al. present a technically-oriented paper on the best practices to create reliable long-term timeseries of atmospheric radon from a two-filter system. This study also includes recommendations and information for atmospheric radon measurements in general as well as a comparison of radon measurements to high-resolution greenhouse gas data. The title does not fully reflect the detailed technical nature of the paper in my opinion.

General comments:

Overall, the paper is well-written, clearly structured and easy to follow. The authors chose great figures to illustrate and communicate their findings and recommendations. The literature cited does however seem a bit arbitrary as there are more studies using radon or the Radon-Tracer-Method. However, as this is not a review paper this is only a minor issue which can be fixed by highlighting that the reference chosen are only a subset of relevant studies. Despite the focus on radon and the technical nature of the manuscript it will definitely be of interest to experts in the field and other readers of AMT, thus I recommend publication after some minor and technical issues have been addressed.

Specific and technical comments:

P2 L 25: It would seem prudent to highlight (make the distinction) that the uncertainty of GHG emissions is mostly an issue for non-CO2 GHGs or carbon cycle feedbacks, while fossil fuel CO2 emissions are typically much better constrained, especially in countries reporting under UNFCCC Annex I

P2 L45: Here and elsewhere:  Suggest to add "for example" before citations to highlight that this is far from a comprehensive list.

A quick googles-scholar search find many more relevant studies on both the use of radon as a transport modelling tracer (e.g. https://doi.org/10.1016/j.jenvrad.2004.03.033, https://doi.org/10.3402/tellusb.v65i0.18681, https://doi.org/10.5194/acp-15-1175-2015) or the use of atmospheric radon and GHG data for the radon tracer method (e.g. https://doi.org/10.5194/acp-7-3737-2007, https://doi.org/10.5194/acp-21-17907-2021, https://doi.org/10.3402/tellusb.v65i0.18037, https://doi.org/10.1080/1943815X.2012.691884).

P3 L53: also many more studies than just Tolk et al.

P3 L58: Worthwhile to cite some papers that use radon as ATM performance tracer.

P3 L69: What is meant by local scale? Urban scale or even facility scale? Yver-Kwok used RTM to estimate emissions from a waste water treatment plant - this was the most 'local' study I was able to find: https://doi.org/10.5194/amtd-6-9181-2013

P9 L200 and elsewhere: here the flow is reported in liters per "m", but before minutes were abbreviated as min (P8 L193). Please make sure to use consistent abbreviations and units throughout the manuscript.

P12 L341: please consider replacing 'networked' with a proper description.

Figure 3: here the internal flow is reported as m3 s-1, while later sections refer to the internal flow in m s-1 (which is odd). Please be sure there is consistent use of units throughout the manuscript.

P14 361f: I assume flow rates are supposed to be "m3 s-1" not "m s-1" here

P14 369: Pressures are given in Pa, while the figure above suggest that the instrument record hPa, why the unnecessary conversion and not put 1-1.2hPa here?

Table 1: same unit issues sometimes per minute is m-1 then min-1 and then m is used for meters in the same table. 

P18 L456: suggest to remove ("to arrive") to clarify the sentence.

P18 L469: Shouldn't this be "radon activity concentration"?

P20 L511: Possion -> Poisson

P23 L577: should be: "atmospheric trace gas constituents". The bulk gas concentrations (N2, O2, Ar) hardly change in the troposphere/ 

P27 L662: Is this new scientific information or is some of this information also in the manual for this ANSTO instrument, if so, it should be referenced.

P27 L673: What is this mobile calibration standard transfer device and where can it be acquired or requested?

P27 L684: Please provide information on the supplier/manufacturer of the Burkert calibration or cite a document that describes it use/function.

P28 L709: was is "production code"?

P29 L721 and section before: This description is helpful, however, the reader is left hanging. Where is all of this data going? Is there a central repository or a global database users can access. If not, is this something you recommend to be created?

P29 L724: This paper only tangentially talks about 'real-time' data. Most of the things discussed here are about high temporal resolution instead. Given the focus on calibrations and the clear week to 5 year time-scale of calibration it seems odd to mention real-time. The data is flagged monthly, calibrated quarterly and deconvolution of the data is done every 6 month (or recommended to be done on that schedule section 5.4), hence reliable data is only available with months delay, i.e. very far from real-time. Also, who would really need/want actually real-time radon data? Reporting data a few hours or even days delayed seems perfectly fine for virtually all applications. 

Table 2A and general: again internal flow is reported in m s-1, but now external flow is L min-1 instead of L m-1 before...