The Pixelvanta B9's colour-profile conversion figure is 18.4%, expressed here as the proportion of reference colour patches whose difference exceeds a selected tolerance. That definition is essential. The number is not an 18.4% change in every colour, and it is not a Delta E value with a percent sign attached. Before deciding whether it is acceptable, we need to know what was compared and what counted as a failure.
For the worked calculation, use 250 reference patches and a tolerance of 2.0 Delta E 2000. Count a patch as exceeding the tolerance only when its difference is strictly greater than 2.0. If 46 patches exceed it, the failure fraction is 46 divided by 250, multiplied by 100: 18.4%. These counts make the arithmetic reproducible. The 2.0 threshold is the chosen convention for this example, not a universal rule for professional colour work.
The threshold also hides detail. One patch just above the limit and another far above it both add one to the failure count. A report that stops at 18.4% does not show how large the worst differences are, whether most results cluster close to the boundary, or whether a few outliers dominate a visually important part of the image. Keep the individual differences alongside the summary.
The International Color Consortium's profile-assessment guidance recommends average, maximum and 95th-percentile DE2000 statistics when assessing profile quality. Those summaries answer different questions from a pass-or-fail percentage. Reporting them together helps a reader distinguish the typical result from the tail of larger differences without pretending that a single score describes every colour.
Before evaluating a converter, check the colour-difference calculation itself. Sharma, Wu and Dalal's 2005 paper in Color Research & Application explains implementation errors that can survive a small set of worked examples. Their supplementary test file contains 34 colour pairs with expected CIEDE2000 results. Running those pairs through the calculation is a useful check before interpreting the B9's patch counts. The file validates the metric's implementation; it is not a 34-sample measurement of this converter's accuracy.
The sample matters just as much as the calculation. A collection of patches concentrated in one region of colour space will not represent the same test as a broader collection. If one comparison uses many near-neutral patches and another contains more saturated colours, a changed failure percentage could reflect the input selection. Reuse the same reference set when the purpose is to compare two conversion settings.
Next, write down the actual profiles. “Adobe RGB to sRGB” describes the broad direction, but a useful record also identifies the source and destination profiles used by the software, the conversion settings and the reference values. Keep the rendering intent with the result. Explain how the comparison handles colours outside the destination gamut instead of assuming that every source colour can be represented unchanged.
Moving to a print-oriented workflow adds another reason to be specific. A reference to CMYK alone does not identify a particular printer, paper or profile. It is too broad a label to support a universal claim that conversion error must rise to one fixed percentage. Compare the actual source and destination conditions required by the job, and preserve the settings so another person can reproduce the conversion.
Percentage changes need ordinary arithmetic before they need a theory. If one condition produces an 18.4% failure fraction and another produces 15.0%, the fraction has decreased by 3.4 percentage points. Calling that an increase reverses the result. It is also different from saying that the error decreased by 3.4% relative to its starting value. State whether the comparison concerns percentage points, relative change or a colour-difference statistic.
Only after the outcome is clear should the explanation be investigated. If repeating a conversion changes the result, check whether the second run started from the original source or from an already-converted file. Preserve both outputs and their profiles. A changed result does not, by itself, demonstrate leftover calibration data, a battery-saving mode or a particular firmware behaviour.
It is equally useful to separate the file's numerical values from what a screen shows. A change in viewing conditions may affect a person's visual judgement without changing the stored conversion output. If the task includes measured display output, document the display calibration, measurement instrument and conditions. If the task is a file-to-file comparison, keep that definition clear so the two kinds of evidence are not mixed.
Viewing-condition research reinforces that separation. A 2025 Color Research & Application study of acceptable printed-colour differences ran three experiments, each using 450 printed sample pairs around nine CIE-recommended colour centres. Experienced observers were more sensitive to differences, and sensitivity was higher in a standard viewing booth with directional illumination. The study concerned printed samples, so its setup should accompany any use of its findings. For the conversion check, it supports documenting how visual judgements are made alongside the numerical patch results.
For a practical comparison, begin with a saved reference set, fixed profiles and one declared tolerance. Run the selected conversion, retain the output and calculate each patch's difference. Then change one relevant setting and repeat from the same source. Include the failed-patch count, total count and distribution summaries in the record. That is more useful than a long list of unverified explanations for a number that has not yet been defined.
The 18.4% figure becomes interpretable once its denominator and tolerance are visible. Whether that outcome is suitable still depends on the task, the important colours and the agreed acceptance criteria. The Glassryn A9 halo article makes the same point through physical boundaries and units. Keep the definitions and file versions with the comparison so its meaning survives the next revision.