Begin with the business property.
Renaming an amount column to revenue does not establish what the number represents. It might be a gross order amount, a net amount after discounts or a paid amount after settlement. The model should state the intended meaning before pointing to the field that supplies it.
In the Ontomato sample, the measure is intentionally narrow: the order amount in USD. The calculation does not label it a financial loss. Each order carries one numeric amount, and the result explains which orders matched. This keeps the claim aligned with the available data.
A useful mapping records the source, field, business property and any transformation. If values are converted between units or currencies, that conversion needs its own rule and context. If a value is missing, preserve the difference between unknown and zero.
Relationships depend on keys.
The simplified dataset uses products.batch to find batches.id, orders.product to find products.id, and feedback.order to find orders.id. Those mappings form the route from a feedback record back to the batch used by the ordered product. A connection in a drawing is not a substitute for these explicit keys.
Before using a mapping, inspect duplicate identifiers, missing references and inconsistent formatting. Decide what happens when a referenced record cannot be resolved. Silently attaching it to a similarly named record would create a relationship the data did not establish.
Cardinality matters too. This teaching sample gives each order one product and each product one batch. A real business may need order lines, multiple component batches and a many-to-many link table. Model that structure explicitly before reusing the calculation.
Keep the mapping with the answer.
An answer is easier to inspect when the user can see both the business label and the source behind it. The demo’s Ontology view shows the model route; the Evidence view identifies the matching feedback. These views explain the result rather than asking the user to trust a polished chart.
For document-based data, the same principle applies to passages: preserve the document version and location supporting an extracted property. Automatic extraction remains a planned product direction here. A designed mapping should still specify how a reviewer would check the original statement.
Start with a small mapping table and a reproducible result. A connector can move data into reach; an explicit mapping is what makes that data meaningful for the question.