Reading is one thing.
Writing is another.
KOSMOS validates live against the data model and the sector’s vertical ontology — the structure of the catalogue, not merely its contents — but never writes to it, not even through a table of its own that would later update it. That is structural, not a preference.
What happens when data does not reconcile
Sooner or later it happens: an attribute the catalogue does not associate with that product type, a brand that does not exist yet, a configurability axis nobody has ever declared. A convenient system creates the missing entry at that moment and moves on. That is precisely how a shared catalogue decays: six months later there are four ways of writing the same thing, every one of them created by a well-meaning import.
KOSMOS does one of the two opposite things. Either it stops, and asks a person to fix the entry where that entry is governed. Or it raises a request into a queue an administrator reviews — the same queue that already exists for missing brands and categories — and links the result back on the next load. In neither case does it write.
That costs a few extra minutes, on one case now and then. In exchange the catalogue stays one thing for everyone reading it, which is the only reason to have a shared one at all.
A shared catalogue works for exactly as long as nobody can quietly widen it.
Where KOSMOS sits in the pipeline
KOSMOS is the first step: it takes raw supplier data and carries it into the sector’s ontology, as a normalized product identity. From there the central catalogue remains the authority on what a product is, and every other platform in the suite — marketplace, commerce, warehouse, B2B — reads that same identity instead of building one of its own.
The rules that are not negotiable
The catalogue is not written from here
KOSMOS reads brands, product types, attributes, categories and the relationships between them live — the ontology, not just the values. It never writes to it, under any circumstances, not even after an approval.
The human beats the machine
Every proposed value is editable, and a person’s edit always takes precedence over a cluster’s or a model’s. No proposal is applied without someone being able to see it and change it.
Whoever proposed it says so
A value suggested by AI carries the mark that says so, and that mark disappears the moment a person replaces it. It is never placed on data read from the file or from the catalogue.
Anagraphic and commercial stay apart
Anagraphic data is universal and validates against the catalogue. Commercial data — stock, prices, discounts — is one client’s own policy and is never inferred from somebody else’s catalogue.
The most useful file is the messiest one.
Send us a price list exactly as it reaches you. We will tell you what would come out of it, and where a person would still be needed.
Let’s talk →