When repeated scans disagree, investigate the differing fields
Use a small repeat-run log to distinguish unstable outputs from verified package values without treating majority agreement as proof.
If repeated scans of the same label give different results, save the outputs and identify exactly which fields changed. Do not keep scanning until one result matches your expectation. Repetition is useful evidence about consistency, but the package remains the reference for transcription.
Hold the input constant first
Use the same saved image for an initial repeat check if the tool permits it. Record the date, app version and relevant settings. If you take a new photograph each time, note that the input changed as well as the result; the comparison then cannot isolate output variation on a fixed image.
Capture the serving heading with the nutrient rows. The FDA serving-size explanation shows why the food amount belongs with the quantities. A repeat result with a different denominator needs investigation even when its nutrient numbers remain unchanged.
Make a field-level log
In a hypothetical three-run check, a label image yields a serving weight of 30 g each time, while a sodium field appears as 90 mg, 900 mg and 90 mg. Save all three observations. The repeated 90 mg value is not automatically correct because it appears twice; inspect the printed row to establish the transcription.
Record missing fields separately from incorrect fields. If a run returns no amount, that is not a numeric zero and should not be averaged with the other outputs. Likewise, a different product match is an identity problem, not merely a fluctuating nutrient value.
Change one capture condition at a time
If the source image is unclear, take a better photograph and label it as a new input. Keep the earlier image and results. Changing lighting, cropping and orientation simultaneously may improve readability, but it does not tell you which change mattered.
Once the package value is readable, create a manually verified record and identify the evidence. Recalculate any comparison that used a different output. If the text remains unreadable, mark the field unresolved instead of using a majority vote.
A useful conclusion is narrow: this field varied across three runs on this input, and the package check established this value. Avoid generalizing that observation into an accuracy rate for the whole service. More packages and a predefined procedure would be needed for a broader evaluation.