1. The observation and the explanation are different records
Imagine opening a public report that places two sky measurements on opposite sides of a vertical line labeled intervention. The second measurement is higher. A caption says the intervention changed the atmosphere. The chart may accurately reproduce both numbers while the caption still goes beyond what they establish. To assess it, a reader needs to know what was measured, whether the observations are comparable, and what would plausibly have happened during that interval without the proposed cause.
Those are separate questions. An instrument can produce a useful record without identifying the source of every change in that record. A comparison can reveal a difference without assigning that difference to an activity. An activity can be documented without demonstrating its atmospheric consequences. Treating these statements separately makes it possible to accept good evidence while declining a stronger claim that the evidence does not yet support. Skepticism need not mean rejecting the measurement itself.
NOAA's Global Monitoring Laboratory describes its aerosol research in terms of means, variability, trends and the factors controlling aerosol properties. Its scientific-background page emphasizes long records for representing climatological variability and for evaluating models and remote-sensing retrievals. These are monitoring objectives, not a claim that every departure from an average identifies a deliberate intervention. The page's historical network-size and trend examples should not be silently repackaged as a current census or a new event. NOAA GML aerosol research and monitoring objectives
This article examines the gap between noticing a change and attributing it. It does not establish that a particular intervention occurred, evaluate a secret program, or offer instructions for releasing material. It also does not turn an aerosol observation into a temperature forecast or a carbon-removal total. Its practical question is narrower: what must survive in a public evidence packet for someone else to understand how an atmospheric claim was assembled?
The worked records below belong to Juniper, a wholly fictional reporting exercise. Every number, station label, date sequence and comparison is hand-authored. A local script checks arithmetic and record joins; it does not retrieve atmospheric measurements, simulate transport, or perform a real experiment. Separate examples use different invented quantities and should not be combined into a description of one physical system. This explicit separation lets the examples test reasoning without inventing a news event.
Juniper's opening folder contains four items: a graph, a table of values, a statement that an activity was planned, and a short explanation of expected effects. None automatically supplies what the others omit. The graph may lack the table's quality flags. The plan may not document execution. The explanation may describe a possible mechanism without measuring its operation. A good review begins by preserving these differences rather than treating four documents as four independent confirmations of the same conclusion.
2. Give every number a measurement identity
NOAA's Chemical Sciences Laboratory lists instruments for different aerosol observables, including size distributions, scattering, absorption, extinction and composition. Its instrument table is useful precisely because these are separate entries. A composition measurement and an optical measurement can contribute to the same investigation without becoming interchangeable quantities. The table does not authorize a universal conversion between them or establish the source of material in a particular sample. NOAA CSL aerosol instruments and measured properties
Juniper's first spreadsheet has a column named aerosol and three rows reading 12, 14 and 0.20. That heading is inadequate. The first row came from an invented instrument-output channel called S; the second from a mass-like reporting field called M; the third from a column-optics product called O. The exercise deliberately supplies no conversion among them. Averaging the three values would produce a number, but not a coherent measurement. More decimal places would make the result look tidier without repairing its meaning.
The repaired table assigns each value a quantity name, unit or dimensionless designation, location identifier, observation interval, processing identifier and stated quality status. It also retains the original field name so that the editor can trace the rewritten description back to its source. A display label is not enough if it erases a distinction present in the underlying record. The relevant question is whether two rows describe the same thing under the same convention, not whether their column headings can be made to match.
Consider two entirely invented files. File A contains channel S at 09:00, with a ten-minute interval ending at that time. File B contains channel S at 09:00, with a ten-minute interval beginning then. Joining on the displayed time creates a visually convincing pair, yet the intervals do not cover the same ten minutes. Correcting the join does not require declaring either file defective. It requires recognizing that the label represented different boundaries. An honest comparison can retain both records and explain why they were not paired.
A second problem appears when one file uses station identifier J1 and another uses the human-readable name Juniper North. A directory could establish that these refer to the same installation during the relevant period. Without that directory, an analyst should not simply assume identity because the names sound related. Conversely, two installations might share a regional display name. Location matching is a documented operation, not an aesthetic judgment about similar labels.
The outcome of this repair is useful even before any causal analysis. An editor can now state that two particular outputs were compared over a defined interval, while the third was retained as a different kind of information. The report becomes less sweeping but more reproducible. If later evidence supplies a defensible relationship between the quantities, that relationship can be added explicitly rather than hidden inside a spreadsheet formula that readers never saw.
3. A total does not preserve its vertical arrangement
The distinction between a column and a profile is visible in actual observing systems. The AERONET Version 3 paper describes column-integrated aerosol optical depth from Sun photometry. NASA's MPLNET Version 3 product documentation separately lists aerosol heights and profiles of extinction, backscatter and depolarization. These products offer different information; a column value alone is not a vertical map or a surface mass concentration. AERONET Version 3 measurement paper NASA MPLNET Version 3 products
For an original bookkeeping illustration, give Juniper two abstract layers whose contributions add to a dimensionless total. In case A, the lower layer contributes 0.05 and the upper layer contributes 0.15. In case B, the lower contribution is 0.15 and the upper contribution is 0.05. Both totals are 0.20. The lower contribution is three times as large in B, while the upper contribution is one third as large. Knowing only the total cannot distinguish these deliberately constructed cases.
This is an addition exercise, not an atmospheric retrieval. The layers are not assigned actual heights, aerosol compositions or human exposure values. Their numbers do not predict how any instrument would respond. The point is the loss of arrangement when two components are compressed into one sum. A later claim about the lower layer needs information that the total did not retain. It cannot be recovered simply by giving the total a more specific caption.
Now suppose Juniper's graphic places the 0.20 total beside a photograph of a street. A reader might reasonably think that the value describes the air at that street. The underlying record, however, carries no street-level quantity. The photograph has supplied an implied location and interpretation that the number did not establish. The correction is not to discard the photograph automatically, but to label what the number represents and avoid using visual proximity as a substitute for a measurement relationship.
A different mistake would be to combine a profile from Tuesday with a total from Friday and call them a single observed state. In the fictional packet, the profile's timestamp is absent from the caption but present in its filename. Restoring that timestamp changes the comparison. The two products may still offer context, but their agreement is not a same-time confirmation. The report should distinguish an explanatory illustration from evidence that two instruments observed the same conditions together.
There is also a useful negative finding here. If both totals remain 0.20, the editor can say that this particular summary did not change between the two constructed cases. The editor cannot say that nothing changed anywhere in either case. Equally, a changed lower contribution would not establish a changed total. Preserving both the result and its information limit is more accurate than choosing whichever description sounds most dramatic.
4. Quality labels belong to a particular product and version
NASA's current AERONET homepage identifies Version 3 aerosol optical-depth levels as 1.0, unscreened; 1.5, cloud-screened and quality-controlled; and 2.0, quality-assured. It notes that dependent products may apply additional checks. Those labels describe the named product family, not a universal ranking that can be transferred unchanged to every environmental dataset. The homepage was checked for this article on October 30, 2026. AERONET's published Version 3 quality-level definitions
MPLNET's own Version 3 documentation adds an important product-specific warning: downstream screening means that filtering a Level 1 product by its own criteria does not generally reproduce the corresponding Level 1.5 or Level 2 product; its stated exception is NRB. It also says Level 1 data should not be used for publication. These are MPLNET instructions, not an assessment of Juniper's invented data. MPLNET processing levels and screening relationships
Juniper's archive contains two fictional releases, called Draft and Reviewed rather than borrowing NASA's labels. Draft has four observations, identified a through d, with values 2, 4, 8 and 10. Reviewed retains a, b and d, with values 3, 5 and 11; c is absent and its exclusion reason is recorded. The mean changes from 6 to about 6.333. Reporting only that small shift would conceal two different changes: retained values changed, and membership changed.
For the three shared identifiers, Draft values are 2, 4 and 10, averaging about 5.333. Reviewed values are 3, 5 and 11, averaging about 6.333. Their paired change is exactly one unit each. The difference between the complete-release means is only one third of a unit because Draft's additional value of eight also affects its average. Neither comparison is arithmetically wrong. They answer different questions about different sets of observations.
The editor should therefore keep a membership reconciliation alongside the headline result. Three records were retained and changed; one was omitted; no new identifiers appeared. That small table explains why a revised average cannot by itself reveal how each observation changed. It also prevents the excluded record from being silently treated as zero. A missing row and a zero-valued row are different objects, even when a plotting program renders both as a gap near the bottom of a chart.
This exercise does not prescribe an atmospheric screening rule or reproduce either NASA processing chain. Its invented release change could represent any documented review process. The practical requirement is narrower: if an article compares two exports, readers should be able to see whether it compared the same records. A release name, retrieval date and list of identifiers can sometimes explain more than another paragraph describing the newer file as better.
5. Separate reprocessing from a change in the atmosphere
Giles and colleagues' Version 3 paper was published on January 11, 2019, after acceptance on December 13, 2018. It describes automated cloud and instrument-anomaly controls and a 2018 reprocessing of the AERONET database. It also distinguishes comparisons of complete monthly averages from time-matched comparisons. This is a dated methods result, not a newly announced 2026 atmospheric change or a guarantee about every individual deployment. Published AERONET Version 3 methods and comparison study
Imagine that Juniper's team downloads an older period on Monday and a newer period on Friday. Between the downloads, its fictional data provider revises the processing. The resulting plot places the older export before the intervention marker and the revised export after it. The plotted break now coincides with two events: the proposed activity and the change in processing. The chart does not isolate their contributions merely because it draws only one vertical line.
A useful repair is to obtain comparable exports under a documented convention, if they are available, and preserve the originals. In this exercise, both periods can be represented under each fictional processing version. Version A gives an earlier mean of 10 and a later mean of 14. Version B gives 11 and 15. Each consistent comparison shows a rise of four. Mixing earlier A with later B produces a rise of five. Mixing earlier B with later A produces a rise of three.
The four combinations are worth showing because they identify a specific source of disagreement. The observed values in the invented table are fixed. The analyst's choice of which releases to combine changes the reported rise. That does not prove that a provider manipulated anything, nor does it establish that the underlying rise has a particular cause. It demonstrates that the comparison needs a version rule before the result is interpreted.
Juniper's editor could record the correction in a short note: the first graphic compared different processing releases; the replacement uses Version B for both periods; its displayed difference is four rather than five. This is better than silently replacing the image. Readers who saved the earlier version can understand why the figures changed. The note need not accuse the original analyst of bad faith or pretend that a routine correction resolves the separate attribution question.
There is a second useful preservation decision. Keep the source files that generated each graphic, not only the final image. A caption can describe the selected convention, but the files establish which rows were used. If a later reviewer finds that an identifier was duplicated, the analysis can be repaired without reconstructing every step from a screenshot. In Juniper, a transparent revision record turns an apparent dispute about the sky into a tractable question about two particular exports.
6. Preserve the sampling conditions instead of inventing a correction
NOAA's GSN instrumentation page describes a specific arrangement with one nephelometer maintained near 50 percent relative humidity and another sampling controlled humidities. It also describes size selection and distinct optical and chemical measurements. This undated station description illustrates why sampling conditions belong in a measurement record; it is not evidence that every NOAA station currently uses that configuration. No numerical humidity correction is inferred here from the page. NOAA's GSN instrument and humidity-control description
Juniper has an invented record with output S equal to 20 under setting R and output S equal to 30 under setting H. An accompanying note says only that the settings differ. The editor initially writes that the atmospheric quantity increased by 50 percent. The arithmetic, 30 divided by 20 minus one, is correct. The interpretation is not established because the comparison mixes settings without a relationship that would make them equivalent.
The archive contains a second pair, both under setting R, with outputs 20 and 22. That pair gives a ten-percent increase in the specified output. It still does not identify a cause, but it avoids the first comparison's particular setting mismatch. The article can report that narrower result and retain the H observation separately. It should not describe the change from 50 to ten percent as a humidity correction: no physical response function was supplied or estimated.
Suppose a colleague proposes simply dividing every H value by 1.5 because the first pair happened to have that ratio. Juniper's example gives no basis for such a rule. It contains one pair, not evidence that the ratio holds at other values, settings or times. Applying it to another H output of 60 would generate 40, but that would be the result of an invented conversion, not a recovered observation. The appropriate entry is unavailable under a common convention, not a confident-looking substitute.
This distinction becomes especially important in public explanations because a repaired chart can look as if it has solved the entire problem. Here the repair only establishes which comparisons are supported by the supplied records. It does not validate the instrument, estimate a missing response curve, or establish that the later R observation represents the same background conditions as the earlier one. Those would require additional evidence rather than a more elaborate caption.
The useful publication artifact is therefore a small mismatch note: the original comparison mixed R and H; a same-setting pair exists and shows a smaller change; the relationship between settings remains unspecified. That note makes the uncertainty actionable. A reader knows exactly what additional documentation would matter, and the analyst has not filled the gap with an arbitrary formula. The result is neither proof of an intervention nor proof that the observations are worthless.
7. A control comparison is an assumption made visible
Now give Juniper two fictional locations, A and B, and one abstract index measured under a consistent convention. Before the proposed activity, A averages 10 and B averages eight. Afterward, A averages 14 and B averages 11. A rose by four; B rose by three. Subtracting those rises gives one. The arithmetic is straightforward. The difficult question is whether B's change represents what would have happened at A in the absence of the proposed activity.
To make that question concrete, write two possible stories that produce exactly the same observed means. In story One, the background at A would have risen by three, and an additional effect of one produced the observed rise of four. In story Two, the background at A would have risen by four, and the additional effect was zero. B rises by three in both stories. Neither story alters any of the four observed numbers. The difference lies in an unobserved condition at A.
These are alternative constructions, not atmospheric predictions. Their purpose is to demonstrate that the four means alone do not select between them. A spreadsheet can verify the subtraction perfectly while leaving this ambiguity untouched. Calling B a control does not add the missing information. The name describes its intended role in the analysis; it does not establish that it actually supplies the relevant comparison.
A third constructed story makes the limitation sharper. Suppose A's background would have risen by five and the activity reduced the index by one. The observed rise at A is still four, while B still rises by three. The same data now coexist with an effect in the opposite direction from the simple difference of changes. This does not mean all three stories are equally plausible in a real setting. It means their relative plausibility requires evidence beyond the four numbers in this exercise.
Juniper next supplies an earlier pair of periods. A previously rose from nine to ten while B rose from seven to eight. Their changes then matched. That observation is relevant to the proposed comparison, but it does not logically force the next changes to match in the absence of an activity. One matching earlier increment is not a guarantee about a later unobserved outcome. A report can describe the historical agreement without converting it into certainty.
A second candidate location, C, changes from six to 11 during the main comparison, a rise of five. Subtracting C's change from A's gives minus one instead of plus one. The choice between B and C now affects the sign. The editor should ask how the comparison location was selected and whether that choice preceded inspection of the outcome. A selection rationale based on relevant evidence is different from choosing whichever subtraction gives the preferred headline.
The resulting public statement can be precise: under the stated use of B as the comparison, the difference of changes is one index unit; using C gives minus one; the packet does not establish which location best represents A's unobserved background. That sentence exposes a substantive limitation without pretending that subtraction is controversial. It invites scrutiny of the comparison evidence, which is where the actual interpretive disagreement resides.
8. The observation calendar can create its own contrast
Juniper's next example concerns membership in time rather than processing version. Imagine four equal-duration slots in an earlier period, with authored values 2, 2, 8 and eight. Their average is five. A later export contains only the first two corresponding slots, again with values two and two. Its available-slot average is two. A headline comparing the two exports would announce a decline of three, even though the values in the matched slots did not change at all.
The later values for the other two slots are unknown in the reporting packet. They are not zero. In one hypothetical completion they are eight and eight, making the later full-period average five. In another completion they are 12 and 12, making that average seven. Both completions leave the available observations untouched. The missing slots prevent the available mean from determining the complete-period mean without further assumptions.
This example is deliberately about records, not a claim that a particular observing network loses data in this pattern. It does not prescribe a method for estimating the missing values. Its practical value is showing the question that a missingness table should answer. Were observations planned but absent, recorded but excluded, or never part of the sampling scheme? A blank cell that fails to distinguish those states leaves the reader unable to reconstruct the comparison.
Time weighting adds another choice. In a separate invented series, value two represents one hour and value eight represents three hours. Averaging the two rows gives five. Weighting by the represented durations gives 6.5, because the total is two plus 24 over four hours. Neither result should be called the average without explaining the intended object. One summarizes the two records equally; the other summarizes their stipulated durations equally.
A timestamp does not by itself reveal duration. A record may describe an instant, a preceding interval, or an aggregate supplied by the provider. Juniper's table therefore carries both beginning and ending fields where an interval is intended. This is an editorial choice for the fictional packet, not a claim about the file format of any named network. It prevents the weighting rule from being guessed from how frequently lines happen to appear in an export.
A further trap arises if an article silently changes the meaning of before and after. Juniper's first draft compares a two-hour window before a marker with a six-hour window afterward. Its second draft calls both values daily averages. That label changes the claim even though none of the numbers changes. Correcting the caption requires restoring the actual windows, not smoothing the graph until the two sides look balanced.
The defensible output is a comparison with an explicit observation calendar: which intervals were eligible, which were represented, which were excluded and which remained unknown. If the calendar differs, the report should show the matched comparison as well as the broader available-data summary when those are meaningful. Their disagreement is information about coverage. It should not be disguised as either definitive atmospheric change or evidence that no change occurred.
9. A modeled route is a different kind of evidence
NOAA's public HYSPLIT limitations page warns that meteorological inputs may not adequately represent near-surface conditions or complex terrain, and that spatial and temporal resolution can matter. The page concerns its described public modeling context; it is not a complete inventory of every capability in every HYSPLIT configuration. Its warning supports a bounded reading of a model product, not a claim that transport modeling is useless. NOAA public HYSPLIT limitations and input considerations
Juniper's packet contains an invented map with a line passing near location D. No HYSPLIT calculation was run for this article; the map exists only as a described teaching object. Its caption says a modeled path was compatible with passage through the vicinity. A second document records that equipment was scheduled to operate at D. A third records an elevated value at A. The three items offer different statements, and none contains a direct observation linking a particular material from D to the sample at A.
Suppose the schedule was canceled but the cancellation notice is absent from the packet. The scheduled activity would not establish execution. Suppose instead that execution is independently documented. That would repair the activity record without automatically identifying the material measured at A. In both cases, the proposed chain needs a clear account of what each document contributes. Simply stacking the documents in chronological order cannot supply every missing connection.
The same discipline applies to a chemical-like label in Juniper's invented sample table. The label is deliberately generic and no real compound is specified. A proposed source could contain material with that label, while another source could too. Matching the label does not establish uniqueness unless the packet supplies a reason it is distinctive in the relevant circumstances. The example does not instruct readers to trace real releases; it shows why an attribution claim needs more than a shared descriptor.
There is a useful middle position between certainty and dismissal. An analyst might say that the records are compatible with a proposed explanation while identifying the missing execution or linkage evidence. That is different from assigning a numerical probability unsupported by an explicit model. Juniper's map, schedule and sample do not come with probabilities that can be multiplied into a confidence score. Three apparently consistent items are not automatically three independent tests.
To make this visible, the editor draws a document relationship rather than another atmospheric map. The schedule and public announcement both derive from the same planning memo, while the map and sample table were produced separately. Counting four documents as four independent confirmations would double-count the planning claim. The corrected record names the common origin. This is a provenance finding about the fictional folder, not an accusation against a real operator or a conclusion about any actual source.
10. An error range needs a specified relationship
The Joint Committee for Guides in Metrology's JCGM 100:2008 explains that significant correlations among inputs must be considered when combining measurement uncertainty. Shared instruments, standards or reference data can create such relationships. The document is the September 2008 guide, not a newly issued atmospheric attribution standard. Its relevance here is limited to the relationship among measurement inputs. JCGM 100:2008, section 5.2 on correlated inputs
For a deliberately bounded arithmetic case, let two reported values be 14 and ten, with each allowed to differ by at most one from its stated value. If either deviation can vary independently within those bounds, the difference can range from two to six. If instead both contain exactly the same additive offset, that offset cancels and the difference remains four. These stipulated ranges are not confidence intervals or measured instrument uncertainties. They illustrate why the relationship between errors matters, not merely their individual sizes.
Juniper's spreadsheet supplies no documentation behind its two annotations. The editor therefore retains both conditional cases and labels them unverified, instead of selecting the outcome that best fits a preferred conclusion.
Even a resolved measurement range would not choose among the alternative background stories in section seven. Those stories differed in an unobserved comparison condition, not in the arithmetic precision of the four supplied means. Juniper could measure its index more precisely and still lack the evidence needed to decide between the proposed causal explanations. This is why an attribution report should not fold every unknown into one unexplained error bar.
A useful editorial layout gives separate space to the observed comparison, the documented measurement limitations and the assumptions supporting the causal interpretation. The reader can then see which conclusion would change if one piece of evidence changed. In Juniper, learning the relationship behind the two annotations would resolve one numerical ambiguity. It would not establish that the planned activity occurred, identify a source, or select the best comparison location. Those remain separate open questions.
11. Write a finding that survives the missing documents
Juniper's folder now supports a more disciplined report than its original before-and-after graphic. It contains comparable records under a named convention, a list of exclusions, and alternative control comparisons. It also contains unresolved gaps: an activity plan without execution evidence, a proposed linkage without a distinctive source record, and a missing explanation for two uncertainty annotations. A conclusion should preserve this mixed state rather than force every item into a single verdict.
Consider three possible opening sentences. The first says an intervention changed the atmosphere. The second says there is no evidence of change. Both overstate what Juniper's constructed folder supplies. The third says that a specified comparison shows a change in an index, while the packet does not establish its attribution to the proposed activity. That sentence accepts the demonstrated result and identifies the particular inference that remains unsupported.
The next paragraph can make the finding testable. Under consistent Version B processing, the main before-and-after difference is four. Using comparison location B yields a difference of changes of one; using C yields minus one. The analysis does not determine which comparison is appropriate for the absent-activity condition at A. These statements remain true even if the reader strongly favors or opposes the proposed activity. Their validity rests on the authored table and its stated calculations.
A subsequent document could change part of the account without overturning everything. Suppose Juniper receives a verified execution log. The report can update the activity status from planned to documented as performed, within the log's stated scope. It need not rewrite the numerical comparisons. Nor should it automatically promote the attribution conclusion. The new record fills one identified gap, which is useful precisely because the earlier report did not confuse it with all the others.
Suppose instead that the team finds a duplicate identifier in an export. That discovery changes the analysis inputs and requires a recalculation of the affected comparison. It does not establish whether the activity occurred. By naming each finding and its supporting records separately, the editor can make a targeted correction. Readers are spared the false choice between treating the entire report as infallible and abandoning every valid observation because one component changed.
This modular account also makes disagreement more productive. A critic can challenge the choice of B, the time-window definition or the treatment of a particular record. Each objection points to something that can be examined. A general claim that the chart proves everything or proves nothing is harder to test because it erases the intermediate decisions. Juniper's revised report is stronger not because it sounds more cautious, but because its individual claims have identifiable evidence boundaries.
12. The public value is in the comparison that can be reconstructed
The most useful deliverable from this exercise is not a dramatic conclusion about an invented sky. It is a compact record of how the conclusion was limited. Juniper's retained packet includes the exact input rows, their measurement identities, the release pairing, the observation calendar and the calculations actually performed. It identifies the planned-activity document as a plan and the modeled-route illustration as a different kind of evidence. Nothing is quietly upgraded by placement in the same folder.
A reader can reproduce the numerical comparisons without believing that the synthetic values describe the world. That distinction is central to the exercise. The script establishes that the stated sums, means and conditional ranges follow from the authored inputs. It does not validate an observing network, test an atmospheric intervention, or produce new environmental evidence. A successful arithmetic check is a narrow result, and naming its scope prevents a demonstration from becoming a fabricated experiment.
In a real public report, the corresponding discipline would let readers ask precise questions. Which records describe the same quantity? Which comparison was fixed before the outcome was examined? What changed between releases? Which part of the explanation is observed, which is modeled and which remains assumed? These questions do not require presuming either hidden misconduct or flawless performance. They require the report to show how its claims are connected to its evidence.
An unexplained anomaly can deserve investigation without already identifying its cause. A documented activity can deserve scrutiny without every nearby measurement becoming evidence of its effects. And a well-supported measurement can remain valuable even when attribution is unresolved. Keeping those possibilities separate is not a way to evade conclusions. It is how a conclusion becomes specific enough to be checked, corrected and trusted within the limits of what was actually established.