Three feeds off one record. Take the attributes Harmonya already maintains, deepen the ones you own, or compute a definition you set across every UPC.
A feed is only useful if you can act on a value you disagree with. Every field arrives with enough around it to check.
One resolved answer per field, not whichever source wrote last. Everything collected for that product is normalized into a single record keyed to the UPC before any field leaves it.
Each value can carry a short written rationale, so a field you disagree with gets checked against the evidence that produced it rather than argued about.
Every attribute names the data elements it was allowed to read. A factual field sees the panel and the pack; a conclusion sees the language shoppers used. You can see which was which.
Which one you want depends on what you already have.
The three feeds differ only in what you bring to them. If you would rather not run the catalog at all, that is a different arrangement.
Hand the catalog over and have Harmonya run the coding, the cleansing and the harmonization against a standing commitment.
Thirty minutes on your categories. How the attributes get built, and the questions your team could put to them.