What do we actually know about the product?

Four documents in one folder

Let us extend the educational example from the first article. The selected product is the A6-20 set: six colours in 20 ml tubes. The alphanumeric codes here are fictional manufacturer product codes.

All documents, dates, characteristics and names in this dossier were invented for the analysis. They are the conditions of an exercise, not information about real products.

This folder presents several different problems. The fabric instructions concern another product. The addition of fabric to the price list is not yet sufficiently supported. The two storage ranges for A6-20 conflict. The date the photograph was received does not establish the batch shown on the box.

Combining these issues into one “poor data quality” score is of little use. Each calls for its own response: exclude the unrelated document, request evidence for the property, investigate the discrepancy or obtain a legible marking.

A similar name does not establish identity

In this example, the question about S4 is resolved at the product-code level. Its cotton instructions concern T6-20. There is no basis for applying the recommendation to A6-20, even if the names are similar, the tubes look alike and the same company issued both documents.

A product code must also be read in the context of its provenance. A seller’s internal code and a manufacturer’s code may identify a product differently. An identical short string in two catalogues does not establish a relationship between the records: we need to know who assigned the code and what it identifies.

Depending on the product, model, size, volume, contents, destination market, revision or batch may matter. Six tubes of 20 ml and twelve of 10 ml both total 120 ml. Their total volumes can be compared, but they cannot be merged into one variant: the number of tubes and the volume of each differ.

GTIN is used for standardised identification of trade items. The GS1 standard considers new products separately from changes to existing ones; its clothing example distinguishes combinations of style, colour and size. An identifier must match the specific product and variant, as also reflected in Google Merchant Center’s GTIN requirements. It helps link records; product properties and the applicability of instructions still need separate checks. If no GTIN has been assigned, the gap must not be filled with an invented number.

After the product codes have been compared, S4 remains in the folder as material about T6-20. It does not support fabric suitability in the A6-20 dossier. The question of fabric use for A6-20 remains open: excluding another product’s instructions does not prove incompatibility.

A newer date does not fix the wrong product code

The S3 price list is newer than the S2 instructions. Automatically selecting the latest value would put both fabric and the +35 °C upper storage limit on the product page. But the price-list update date does not explain where those fields came from.

Prices may have been added while old descriptions were retained. The supplier may have received new manufacturer recommendations. A row may have been copied incorrectly. Without further information, all three explanations remain hypotheses.

It is useful to distinguish a document’s issue date, edition, effective period, receipt date and internal review date. “Checked today” means that someone reviewed the record today. It does not turn the contents of 2025 instructions into a new manufacturer recommendation.

Nor should an official document be selected automatically. It may concern a different market, formulation or earlier shipment. To establish priority, check the source’s applicability and purpose: what it establishes, for which product and for what period.

Our question to the supplier should be precise: “For A6-20, instructions S2 give a storage range of +5…+25 °C, while your row S3 gives +5…+35 °C. Which document and edition support the second value? Which shipments does it apply to?” “The price list is current” does not explain the discrepancy.

Store a number together with what it measures

During extraction, “20 ml” looks like a ready-to-use specification. But it is the volume of one tube, not the set. If only the number and unit are retained, a later operation may put it into the wrong field.

A working record needs the measured object, property, value, unit and source. For temperature, it also needs the context: storage, transport or use. Two different values may be compatible because they concern different contexts. That distinction must come from the documents, rather than guesswork.

During normalisation, “0.02 l” can be converted to “20 ml”, retaining the original notation and conversion operation. The 120 ml volume for six such tubes is a separate derived value. It must remain clear which information is document text, which is unit conversion and which is a calculation.

Ranges, signs and approximate values need particular care. “Up to 30 minutes”, “30 minutes” and “about 30 minutes” define different answers. If those distinctions are removed when preparing a table, the description writer will no longer be able to recover them.

Two ranges that cannot be averaged

S2 gives the permitted storage range for A6-20 as +5 to +25 °C, and S3 as +5 to +35 °C. Both records specify complete permitted ranges for the same product code. Their upper limits differ, and the supplied documents do not explain why.

An average of +30 °C would be a new recommendation found in neither source. Choosing the narrower range does not establish which record applies to the current product either. That choice may be a separate handling decision, but it cannot be presented as the manufacturer’s verified recommendation.

Nor should +25 °C simply be retained without comment while all other information is declared reliable. The identified discrepancy must remain in the dossier and affect approval of the corresponding claim.

Other information does not necessarily become invalid. The contents shown on the packaging can be checked independently of the temperature range. However, if storage conditions are mandatory for the material or publication channel, the conflict must be resolved before publication. Removing a mandatory warning does not solve the problem.

Unknown information, a prohibition and a conflict require different actions. Missing fabric information requires a source. A direct prohibition in applicable instructions, if one existed, would need to be retained as a restriction. Two incompatible values require investigation of the discrepancy. Reducing all three states to a blank cell loses the task that still needs resolving.

What a photograph and a review establish

S1 lets us read the contents, but not the batch number. Enlarging the image helps; filling in illegible characters from a similar catalogue code is already an assumption. Recognition output should be checked against the original, especially when a character determines the product variant.

A legible product code establishes which variant the photograph shows. Linking the image to the current shipment requires a separate provenance check: a shared media folder may contain photographs of old packaging. Until this link is established, “the current shipment contains six tubes” goes beyond the available evidence.

Suppose a review is added to the folder: a buyer says they painted on a T-shirt. This is a useful signal for researching a need. It does not establish recommended fabrics, material preparation, fixing conditions or wash resistance. The questions to ask the manufacturer have become more specific; the technical specification has not gained new facts.

An in-house test also needs clear boundaries: which product and sample were used, what was done and what was observed. One painted sample does not establish suitability for all fabrics. An observation can be retained as an observation without quietly being promoted to general instructions.

The history of a value

After extraction and calculation, the editor sees “120 ml”. To check it, they need to trace it back to the packaging: six tubes, 20 ml in each, multiplication. If only the final number remains, it becomes difficult to distinguish a calculation from a direct statement.

This history is called data provenance. The W3C PROV family provides ways to describe entities, activities and agents involved in creating and processing data. In a product dossier, this approach can link a document, an extracted value, a transformation and a published phrase. The chain supports review, but does not itself prove that the original information is correct.

In practice, it is useful to be able to follow this path in both directions. From a finished sentence to its evidence. From an updated source to the dependent values and texts that need review.

If the volume of one tube changes, the total volume, a comparison with another set and an image caption all need review. If only a typo in the filename is corrected, the numerical conclusions do not change. Updates require assessment of their content; otherwise, a system will either miss changes or unnecessarily rewrite the entire catalogue.

Even several identical values do not necessarily provide independent confirmation. Shops may copy the same supplier listing. Five matching pages would then reproduce one original record, including any error. Relationships between sources matter when assessing evidence.

When the dossier is ready for the writer

The A6-20 analysis leaves us with records in different states. The contents were read from the packaging; suitability for primed canvas is stated in the instructions for this product code. The photograph does not establish the batch. Fabric use requires confirmation. The storage range is in conflict.

The writer needs access to these distinctions. If given a single paragraph — “six tubes, 120 ml, for canvas and fabric, storage up to +35 °C” — they cannot see which information is supported, calculated or disputed. Values must be accompanied by essential conditions, status and evidence; disputed information must be excluded from automatic use until a decision is made.

A specific test can help assess a service here. Give it the four documents from our dossier and check whether it can explain why S4 does not support fabric use for A6-20, where 120 ml came from and which conflict prevents the choice of a temperature. Then correct one source document and see which dependent records are flagged for review. Such a task reveals the processing mechanics much more precisely than a general promise to “consider all characteristics”.

Collection can be considered sufficient for a particular task when the information essential to it has applicable supporting evidence and open questions are not hidden. Comparing total volumes may require only the number of tubes and the contents volume of each. Our material is not yet sufficient for storage instructions.

We still need to determine which tasks the buyer is trying to accomplish. They may be interested in ready-made colours, suitability for a particular surface or brush care. A checked dossier limits the answers we can give; deciding what to explain first on a product page starts with studying the questions themselves.