Why Argo tables need translating
An Argo float profile can look like a tidy list of pressure, temperature and salinity. Yet the most important question is often not what the number says, but which number you should use. The raw value, adjusted value, quality flag and uncertainty can tell different parts of the same scientific story. This guide translates Argo float files into decisions that a careful reader can defend, rather than treating a convenient column as unquestionable truth.
Argo float data modes and quality-control flags are especially important when comparing salinity across years, estimating subtle temperature or density changes, or building a regional time series. A value may be plausible and still carry a warning about sensor drift, estimated position, an uncorrected measurement or a calibration not yet completed. Recognising those conditions is a practical reading skill, not merely a software convention.
We will begin with the mechanics of an autonomous profiling float, open a conceptual netCDF profile, choose between raw and adjusted fields, perform checked calculations and test the decisions against realistic failure cases. Every numerical exercise below is explicitly illustrative, not a claim about a particular float or ocean location. The goal is to make the data provenance as readable as the measurements.
Start with what an Argo float actually does
A conventional profiling float drifts below the surface, changes buoyancy to travel through the water column, records properties with its sensors and eventually returns toward the surface to communicate. The resulting record is not a permanently fixed monitoring station. It is a succession of profiles with positions and times that may differ from one another. This matters because two profiles with different temperatures may have sampled different water masses rather than revealing warming of one piece of water.
Pressure, often labelled PRES, is an instrument measurement commonly supplied in decibar. Temperature, TEMP, is reported in degrees Celsius according to file conventions. Practical salinity, PSAL, is a dimensionless numerical quantity on the Practical Salinity Scale; calling it grams per kilogram without making the necessary distinction can introduce a conceptual error. Pressure is not an identical synonym for geometric depth. Near the sea surface one decibar corresponds roughly to one metre, but a precise depth conversion requires latitude and seawater properties.
A float cycle and a sampled vertical level have different roles. A profile is a collection of observations along an ascent or descent; a level is one measurement position inside it. A missing level does not necessarily invalidate every other level. Conversely, a whole profile can have a positioning or timing issue even if a few sensor samples look smooth. Scientific filtering should state whether it removes individual points, complete profiles or entire floats.
Think of each observation as a four-part statement: measurement, where-and-when context, calibration state and uncertainty. Keeping those parts together prevents an easy but serious error: exporting only PRES, TEMP and PSAL to a spreadsheet, then forgetting the flags that explain which values are trustworthy.
Understand the file before interpreting the numbers
Official Argo data are distributed through Global Data Assembly Centres in netCDF files. Filenames and metadata distinguish the data stream, the float identifier and the cycle. Real-time profile files are generally identifiable by an R prefix, and delayed-mode profile files by a D prefix. Core profiles carry conductivity-temperature-depth information; biogeochemical profiles contain other measured variables and may use separate or synthetic profile files to align differing sampling patterns.
A netCDF file is more like a labelled multidimensional workbook than a simple comma-separated table. Dimensions can represent number of profiles, number of levels, number of parameters and fixed-length character fields. A program should read the declared dimensions and missing-value attributes rather than assuming that every file has one profile or that every array is aligned by row number. Even where an example file is simple, that habit makes the interpretation portable.
Key metadata include latitude, longitude, time, cycle number, direction, data mode and the parameter list. Names such as JULD, LATITUDE and LONGITUDE locate the observation, while corresponding QC fields warn about estimated or questionable coordinates. Under sea ice, a float may be unable to obtain a direct satellite position for an interval. Treating an estimated position as a precise fix can make an attractive map misleading.
A defensible export carries the source file identifier, download or update date, parameter name, chosen raw-or-adjusted field, associated QC flag and stated error. This may feel like extra bookkeeping until another analyst asks why your regional mean differs from theirs. With the trail intact, the discrepancy becomes an answerable question instead of a mystery.
Read R, A and D as processing states, not grades
Core Argo profile data use DATA_MODE to indicate whether a given record should be read from raw or adjusted variables. In R mode, adjusted data are not available for that selection and the raw parameter is the relevant field. In A mode, real-time adjustments are available. In D mode, delayed-mode calibration and review have been applied. The important distinction is the source of the scientifically selected value, not an informal ranking where a letter magically certifies every measurement.
For a core profile, the Argo guide says to use PARAM when the mode is R, and PARAM_ADJUSTED when the mode is A or D. The corresponding QC field should be selected consistently. It would be contradictory to use PSAL_ADJUSTED but attach the quality flag from raw PSAL as though both referred to exactly the same evaluated value. Write that selection rule in code rather than relying on memory.
Biogeochemical Argo has an additional complication. A single float may have oxygen, nitrate and optical signals at different stages of adjustment. The overall DATA_MODE is not sufficient for choosing each biogeochemical parameter. Examine PARAMETER_DATA_MODE for the relevant parameter. The good habit is to ask ‘which parameter, in which profile, in which mode?’ rather than ‘is this float a delayed-mode float?’
Delayed mode is a review process, not time travel. Scientists may revisit calibration when a longer sequence of measurements or better reference observations becomes available. A previously downloaded file can therefore differ from a newly distributed revision. For a reproducible analysis, record both the dataset version or retrieval context and the selection rules. Never silently replace values halfway through a long analysis.
What the quality flags actually permit
Argo QC flags have named meanings. Flag 1 denotes good data; 2 means probably good. Flag 3 indicates bad data that may be correctable, while 4 marks bad data. A zero indicates that no quality control was performed. Eight represents an estimated value and nine a missing one. Other flag codes have specialised meanings, so a generic test that assumes every numeric value below four is acceptable would be indefensible.
The correct rule depends on the scientific claim. For a stringent estimate of subtle salinity changes, choosing delayed-mode adjusted values with adjusted QC equal to 1 is a sensible starting point. For an exploratory map in a sparsely sampled region, an analyst might retain probably good values with an explicit caveat and sensitivity check. What matters is declaring the policy before looking at whether the resulting picture is aesthetically pleasing.
Treat flagged numbers as evidence about measurement status, not as inconvenient outliers to be smoothed away. A smooth profile could be a drifting sensor; a sharp feature could be a genuine front. Automatic spike tests help but do not substitute for examining temperature, pressure, geography and neighboring observations. Rejecting every sharp feature risks deleting the event that the study intended to detect.
Missing data should remain missing in the primary analysis. Interpolation can be appropriate for a clearly described purpose, but the interpolated number must not masquerade as a direct observation. In particular, a gap across a strong thermocline cannot safely be filled by a straight line and then used as if it resolved the fine vertical structure.
Adjusted values and uncertainties belong together
An adjusted observation is not automatically exact. Its companion field, such as PSAL_ADJUSTED_ERROR, records information about estimated uncertainty. Because the adjustment may account for sensor drift, reference-data mismatch and other effects, the uncertainty is part of the result. A table that retains a corrected value but drops the corrected error communicates unjustified precision.
Suppose a raw practical salinity value is 35.018, while an illustrative delayed-mode value is 35.006 and its associated adjustment uncertainty is 0.004. The adjustment is minus 0.012 in practical salinity units. It would be wrong to describe 35.006 as the exact true salinity, and equally wrong to claim that the minus 0.012 correction proves environmental freshening. This is one measurement whose estimate has been recalibrated.
Uncertainty is not always a single additive number. Random sampling scatter, calibration uncertainty, spatial mismatch and interpolation uncertainty describe different mechanisms. If two independent estimates with standard uncertainties 0.004 and 0.003 are subtracted, the standard uncertainty of their difference is the square root of 0.004 squared plus 0.003 squared, or 0.005. If errors are correlated because the same calibration influenced both, that formula needs a covariance term.
For useful conclusions, report the size of the observed difference beside an uncertainty assessment and the nature of the comparison. A value differing by 0.002 between profiles may be scientifically real under excellent conditions, but it is not demonstrably resolved merely because three decimal places appear on the screen.
Why salinity drift and thermal lag complicate reading
Conductivity is converted to practical salinity using temperature and pressure information, so a fault in one sensor or an imperfect correction can affect the derived salinity. Calibration can drift over a float’s lifetime. A time series that slowly rises can therefore mix actual environmental change with a gradual instrument offset. The delayed-mode process uses suitable references and expert judgement to distinguish them as well as the evidence permits.
Thermal-mass effects in a conductivity cell can also matter where water temperature changes quickly with depth. If the conductivity cell and the temperature sensor respond differently across a strong gradient, salinity can acquire an artificial feature. Inspecting a narrow salinity spike without temperature, ascent direction or instrument context is poor diagnosis. A responsible interpretation asks whether the shape repeats across profiles and whether known corrections have been applied.
Reference-based corrections are statistical comparisons, not omniscient replacements for shipboard observations at exactly the same place and time. Reference data may be sparse, and natural seasonal or regional variability can look like instrument drift. A good analyst reads the delayed-mode adjustment and its error rather than assuming a correction algorithm has isolated all causes perfectly.
Refrain from inventing a fixed correction tolerance that every ocean project should apply. The acceptable threshold depends on the intended variable, basin, signal size, time scale and published calibration guidance. A broad educational map and a deep-ocean climate trend require different levels of conservatism.
A checked workflow from file to answer
Begin by naming the question. ‘What was the upper-ocean structure in this region last month?’ differs from ‘Has deep salinity changed by two thousandths over a decade?’ The first may tolerate cautious near-real-time data; the second generally demands a delayed-mode review, a consistent sample definition and uncertainty accounting. Decide those requirements before computing an average.
Second, identify the exact file family and parameter. Inspect the metadata, confirm the reported units, isolate the cycle and examine both data mode and QC flags. Third, select the appropriate raw or adjusted field and matching quality indicator. Fourth, reject fill values and invalid observations before doing arithmetic. Fifth, check positions, times and pressure order; do not let a missing timestamp silently become midnight.
Sixth, decide whether to compare observations at common pressure levels. If you interpolate, record the method and flag stretches of the column with broad gaps or rapid gradients. Seventh, calculate derived quantities only after the input data pass the declared policy. Density calculations, for example, need a suitable seawater thermodynamic method and more than salinity alone.
Finally, preserve a small audit trail: source, mode policy, flag policy, interpolation rule, unit conversions, exclusions, uncertainty treatment and code version. A conclusion is much more persuasive when another reader can reconstruct why one profile contributed and another did not.
Worked example: choose the correct field
Imagine a core profile with PSAL = 35.021, PSAL_ADJUSTED = 35.008, DATA_MODE = D, PSAL_QC = 1, PSAL_ADJUSTED_QC = 1 and PSAL_ADJUSTED_ERROR = 0.003. The scientifically selected salinity for a delayed-mode analysis is 35.008, and its associated adjusted QC and error travel with it. The difference, 35.008 minus 35.021, equals minus 0.013. That is a calibration difference, not an observed environmental trend.
Now suppose the same pattern occurs in a second file except DATA_MODE = R and the adjusted field is filled with a missing-value code. For a real-time exploratory display, select PSAL and its raw QC and identify the result as unadjusted. Do not copy an adjusted value from a previous cycle into this row merely to remove a missing cell. A profile-specific selection matters more than the convenience of a rectangular spreadsheet.
A third scenario is a BGC file with PARAMETER_DATA_MODE = D for oxygen but R for nitrate. It is unsafe to write a function that chooses every adjusted BGC variable because the oxygen field is delayed mode. Your data selector must index the mode associated with each parameter. A nitrate reading without an appropriate adjustment may require a stricter use restriction even if the oxygen reading is ready for quantitative analysis.
Worked example: differences and uncertainty
Consider two hypothetical good, delayed-mode salinity estimates at the same pressure: 34.997 with standard uncertainty 0.003 and 35.009 with standard uncertainty 0.004. The second minus the first is +0.012 practical salinity units. Under the simplifying assumption of independent uncertainty, the difference has standard uncertainty sqrt(0.003² + 0.004²) = 0.005. The difference is 2.4 times that standard uncertainty. This alone does not establish a regional trend because water-mass displacement and sampling design also matter.
Change the second uncertainty to 0.010 while keeping the observed numbers the same. The difference’s uncertainty becomes sqrt(0.003² + 0.010²), approximately 0.0104. The numerical increase is unchanged, yet the evidence of a resolved difference is much weaker. A plot that only shows a colored line would conceal the distinction; a useful interpretation reports both the estimate and its evidential strength.
Suppose the two profiles are 250 km apart. Even perfectly calibrated salinity measurements could then sample different circulation features. The uncertainty budget needs a sampling or representativeness component, not merely instrument errors. The correct statement may be ‘these samples differ’ rather than ‘the ocean at a fixed location became saltier’. Spatial and temporal comparability are separate questions.
Worked example: binning a vertical profile
Assume a toy profile has valid salinity values 34.80 at 5 dbar, 34.86 at 15 dbar and 34.98 at 35 dbar. A linear interpolation to 25 dbar between the last two levels gives 34.86 + (25−15)/(35−15) × (34.98−34.86) = 34.92. The calculation is correct for linear interpolation. It does not prove that an unmeasured sharp halocline is absent at 24 dbar.
Now replace the 35-dbar point’s QC flag with 4. A policy requiring good data removes it. Extrapolating to 25 dbar from the two shallow observations because a downstream chart insists on a value would manufacture evidence. The defensible result is missing at 25 dbar under the declared procedure, or an explicitly modelled estimate with a different label and uncertainty.
Pressure bins should not be confused with physical layers of equal thickness unless depth conversion and layer boundaries have been handled properly. A simple arithmetic mean of salinity at irregular pressure levels is not automatically a thickness-weighted layer mean. Explain the weighting and coverage used whenever profiles are summarized.
Worked example: an apparent trend that fails
Imagine three annual estimates at a nominal deep pressure: 34.991, 34.998 and 35.006. The slope from first to last is (35.006−34.991)/2 = 0.0075 practical salinity units per year. Before interpreting this as change, inspect the modes. If the first record is unadjusted, the second has probably-good QC and only the third is delayed-mode adjusted, the series combines differently calibrated inputs.
A repaired comparison might use delayed-mode values once available, the same pressure reference, matching regional coverage and a regression with uncertainty. If those conditions cannot be met, keep the initial result as a descriptive calculation and label the trend unsupported. ‘Interesting but not yet resolved’ is an accurate outcome, not a failed article.
Even a fully corrected three-point series is short. It is sensitive to year-to-year physical variability, changes in sampling latitude and the way seasons align. A meaningful climate claim needs a larger independent evidence base and treatment of autocorrelation. A clear method can show a striking slope while refusing to exaggerate what the evidence supports.
What to check when building maps and sections
Profile positions should represent where the data were actually collected or the best available estimated location, not where a trajectory happens to be drawn between surfacing events. Check POSITION_QC and time flags. A float may drift substantially over the interval between communication events; a precise-looking line between fixes is a reconstruction, not continuous GPS observations.
For a vertical section, decide how to group observations across varying time and space. A section assembled over weeks should not be captioned as an instantaneous wall of water. Map sampling gaps explicitly. Interpolation across an unobserved boundary can create a smooth ocean that never existed, or obscure a front that matters to the user’s question.
Depth, density surfaces and pressure surfaces are different comparison coordinates. A salinity feature can move vertically as a water mass heaves; following a fixed pressure may show change that differs from following the same density surface. The correct coordinate depends on whether the user wants fixed-depth conditions, water-mass properties or transport structure.
Combine Argo with independent instruments only after reconciling scales and quality decisions. Our companion guide on ocean CTD, salinity and density discusses how pointwise water-column measurements differ from broad conclusions. The companion on drifter tracks and currents helps distinguish a float’s location from a direct measure of transport.
A practical decision tree for common use cases
For operational situational awareness, start with the newest available good-quality observations, disclose their near-real-time state and compare with other independent measurements. Avoid pretending that quick delivery and final calibration are synonymous. In a rapidly evolving ocean event, timeliness may legitimately matter more than the smallest possible calibration error, but that tradeoff should be visible.
For a classroom graph, a short, openly fictional dataset can demonstrate raw-versus-adjusted selection and flags without implying that the values came from a named cruise. For a regional synthesis, choose a consistent policy for filtering, grid spacing, time intervals and spatial coverage. For a small climate signal, narrow the dataset toward expert-reviewed delayed-mode observations and propagate or display the relevant errors.
When a study becomes sensitive to thousandths of salinity, systematic calibration, representativeness and spatial matching can dominate the story. If a reviewer questions the result, the answer must not be ‘the chart looked smooth’. It should describe the modes used, filters applied, comparisons made and uncertainty left unresolved.
Errors that are easy to make, and concrete repairs
Error one is to treat DATA_MODE as a personal judgement of whether a float is good. Repair it by selecting the parameter field according to the documented mode, then checking the separate QC indicators and errors. Error two is to attach raw QC to adjusted values. Repair it by pairing every chosen value with the corresponding flag and estimate of uncertainty.
Error three is to treat an empty adjusted field as numerical zero, which can create fictitious low-salinity events. Repair it by applying the netCDF missing-value convention before any arithmetic. Error four is to average all profile levels equally and call the result a layer mean. Repair it by defining valid vertical coverage and physically appropriate weighting.
Error five is to infer a stationary time series from a drifting network. Repair it by plotting the locations, seasons and pressure positions, and testing the conclusion against the shift in sampling. Error six is to regard QC flag 2 as interchangeable with QC flag 1 without stating an inclusion policy. Repair it by presenting a primary strict result and, if useful, a sensitivity result that includes probably-good observations.
Error seven is to suppose a corrected salinity is ground truth. Repair it by retaining the published adjustment error and considering reference uncertainty. Error eight is to show interpolated, estimated or geolocated positions with the same visual confidence as directly measured locations. Repair it by distinguishing observational from reconstructed points in the legend and caption.
Error nine is to forget that a downloaded collection may be revised. Repair it by preserving retrieval information and file versions. Error ten is to interpret a large amount of high-resolution data as a large amount of independent evidence. Closely spaced samples within a profile are related; adjacent cycles and nearby floats may also be correlated. An effective sample size can be smaller than the row count.
Exercises for independent transfer
Exercise 1. A core file reports DATA_MODE = A, TEMP = 12.41 and TEMP_ADJUSTED = 12.39 with adjusted QC 1. Which field belongs in an analysis using the best available calibrated value? Answer: 12.39 from TEMP_ADJUSTED, with its adjusted QC and accompanying error field. The A indicates adjusted data are available, not that every other parameter necessarily has no remaining uncertainty.
Exercise 2. DATA_MODE = R, PSAL = 34.72 and PSAL_ADJUSTED is missing. Can the blank be filled with 34.72 and relabelled adjusted? No. The value can be analysed as raw within a declared real-time use case if it passes the relevant QC, but copying it to an adjusted field would misrepresent the processing history.
Exercise 3. A selected adjusted salinity is 35.010 while the raw observation was 35.026. Compute the correction. Answer: 35.010−35.026 = −0.016 practical salinity units. The negative sign means the adjusted estimate is lower, not that the ocean itself freshened by that amount between two dates.
Exercise 4. Two independent estimates have standard uncertainties 0.006 and 0.008. What is the standard uncertainty of their difference? Answer: sqrt(0.006²+0.008²)=0.010. If the estimates share a common calibration error, independence fails; a covariance-aware calculation is then required.
Exercise 5. A measured value carries QC flag 4 but sits exactly on a smooth regional curve. Can the visual fit override the flag? Not by itself. Keep the sample excluded under a strict protocol and investigate the data provider’s record. Agreement with neighboring points is evidence for inspection, not permission to discard the documented warning.
Exercise 6. A vertical profile has values at 10 and 40 dbar but none between. You need a 20-dbar temperature for a teaching illustration. A linear interpolated value may be shown if its derivation and limitations are explicit. It must not be presented as a direct sensor observation or as proof that no thermocline exists between levels.
Exercise 7. A float’s GPS fix is obtained after surfacing, but its profile was recorded over a prior underwater interval. Is the surface coordinate an exact location for every level? No. It is part of the trajectory context, and the file’s position metadata and QC indicators should guide the spatial treatment. Resolve the distinction before estimating coastal gradients.
Exercise 8. In a BGC file, oxygen has parameter mode D and nitrate has parameter mode R. Should both use adjusted variables? No. Select each according to its own PARAMETER_DATA_MODE and scientific suitability. For demanding nitrate research, a raw nitrate measurement may require exclusion rather than a convenient substitute.
Exercise 9. Three nearby profiles from one float have 150 points each. Is the sample size 450 independent ocean realizations? No. Those points have shared instruments and related location-time structure. Use an aggregation design that respects profile and float dependence, and communicate effective coverage rather than treating every cell as independent.
Exercise 10. A trend changes sign when you exclude QC flag 2 values. What should the report say? The conclusion is sensitive to the inclusion policy. Give both defensible analyses, explain why one is primary, and refrain from claiming a robust trend that vanishes under a reasonable quality filter.
Exercise 11. An adjusted estimate is 34.998 with uncertainty 0.009; a nearby older adjusted observation is 35.003 with uncertainty 0.002. Is the observed difference of −0.005 a resolved freshening? Under an independent-error approximation, the uncertainty of the difference is about 0.0092, larger than the apparent change. Location and temporal sampling could add further ambiguity.
Exercise 12. A map places all float profiles at their first deployment positions. What is the likely error? The network is drifting, so later observations are mislocated. Rebuild using profile-specific coordinates and their quality flags, and show uncertainty where direct fixes are absent.
Exercise 13. Someone says ‘flag 0 is better than flag 1 because zero means no error’. Explain the mistake. Flag 0 means no QC performed; it is a categorical workflow code, not a numerical error magnitude. A reader should consult the documented mapping rather than rank flags as ordinary numbers.
Exercise 14. The pressure array is descending in one file and ascending in another. Should a gradient be computed using row index alone? No. Check sampling direction and actual pressure spacing, then calculate differences using pressure units. Sorting for a derived grid may be appropriate, but preserve the original chronology when interpreting instrument responses.
Exercise 15. A salinity diagram has an unusually sharp near-surface spike. Which evidence should be checked? Inspect temperature gradient, ascent/descent direction, adjacent cycles, relevant QC flags, thermal-mass correction state and local ocean dynamics. Do not presume either a sensor artifact or a real front before examining those independent clues.
Exercise 16. A collaborator reports ‘the ocean warmed 0.08 degrees’ after comparing two different floats in different seasons. What must be repaired? Define a matched space-time-pressure comparison, quantify uncertainty and distinguish water-mass differences from change at a fixed location. The arithmetic difference might be correct while the causal description is unsupported.
Exercise 17. An analyst’s CSV lacks any record of which Argo file version supplied an adjusted salinity. What information should be restored? At minimum float/cycle identity, data source, retrieval time, relevant modes, QC selection, chosen variable and adjustment error. Without these, reproducibility is weakened even if the values happen to be plausible.
Exercise 18. A report mixes good delayed-mode deep observations with near-real-time near-surface observations. Can that be useful? Sometimes, if the analysis requires different timeliness and precision by depth and declares the selection policy clearly. It should not be marketed as a uniformly delayed-mode dataset or used to make unqualified tiny-signal claims.
Frequently asked questions
Is an R file wrong? No. It is a near-real-time file whose limitations are different from a fully reviewed delayed-mode product. It can support appropriate operational and exploratory tasks when used with its flags and uncertainty limits. It is not the preferred starting point for very small long-term signals that require expert calibration.
Does QC flag 1 guarantee a physically correct interpretation? No. The flag concerns data quality as assessed by the provider’s tests and review. It cannot establish that two profiles sampled the same water mass or that a feature was caused by climate, a storm or circulation change. Measurement validity and scientific interpretation are separate layers.
Can pressure in dbar be plotted as metres of depth? Only as an approximation if clearly labelled and acceptable for the use. Pressure to depth conversion is not exact one-to-one and varies with latitude and seawater properties. A precise depth axis needs a suitable conversion.
Can I average Practical Salinity values without converting them to grams per kilogram? You can compute a descriptive numerical average of quality-selected practical salinity values if you state what was averaged. But practical salinity is not the same variable as Absolute Salinity, and thermodynamic calculations require appropriate conventions and transformations.
Why might a delayed-mode value change between downloads? Providers review sensor behavior, reference comparisons and corrections as evidence accumulates. Version changes or updated files can legitimately revise adjusted variables, QC flags or error estimates. Preserve the original dataset identifier and retrieve updated documentation rather than quietly blending versions.
Are biogeochemical floats handled exactly like core floats? Not entirely. Different parameter sensors may be adjusted on different schedules, so PARAMETER_DATA_MODE matters in biogeochemical files. Read parameter-specific metadata and guidance, particularly before using nitrate, oxygen or optical observations in quantitative ecological claims.
What if an interpolated profile looks smoother than the measured points? Smoothness is an output of processing, not proof of greater accuracy. Show measured levels, account for gaps and avoid treating resolution introduced by gridding as new observational information.
What is the fastest safety check? Print the chosen field name, DATA_MODE or PARAMETER_DATA_MODE, QC flag, adjustment error and a few source identifiers beside any plotted statistic. That tiny accompanying table often exposes a category error before it becomes an attractive but unreliable result.
What the evidence supports—and what it does not
Argo makes an enormous part of the ocean more observable, but a float network does not remove the need for careful comparison. Its autonomous instruments visit moving regions at varying intervals and share identifiable calibration histories. Good science makes the observational power visible while refusing to claim more spatial or temporal certainty than the data can carry.
The value of translating a technical file is that a researcher, student or policy reader can move from ‘this column is 35.01’ to ‘this is a delayed-mode calibrated salinity estimate at this pressure, in this profile, with a documented flag and error, compared under a specified sampling rule’. That second sentence has enough structure to be checked. The habit transfers to groundwater sensors, field probes, laboratory instruments and other scientific measurements.
For a related practice in distinguishing measurement from interpretation, see ocean oxygen optode calibration and ocean nutrient quality control. The same principle appears in language: a label carries meaning only when its context and conditions are understood.
Primary references and methodological reading
Argo: How to use Argo profile files
Google Search Central: Creating helpful, reliable, people-first content