Why this screen exists The canonical model is the fixed point. It is the requirement of the AI use case and it cannot be bent per customer, or there is no reusable industry model left to sell. So the customer’s data is what bends, and it bends here. Handing that work back to the client is where these projects stall.
8 fields need one, 0 run ← back to Mapping Every one of these came out of an accepted mapping whose shapes did not line up. A transform is judged on the rows it fails, not the ones it handles.

Where a transform runs

Pushdown first. Most of these are a CASE, a split_part or a cast, and the source can do them without a row moving. Platform compute is for what a source cannot express or does not hold.

pushdown Compiled to SQL and run in the source. No data moves, no store to secure, no schedule to miss. 6 of 8
platform Run by the layer. Needed when a value came from a document, or when the join crosses two sources. 2 of 8

Pushdown is the difference between owning a compute cluster per customer and owning a SQL compiler. Only the two that genuinely cannot push down are run here.

What a transform is not

Reshaping something the business recorded is a transform. Producing something it never recorded is inference, and it gets a different word on this screen and a different treatment downstream.

ict_asset_inventory.dependencies is required and nothing at Nordbank holds it. It can be approximated from incidents that took several assets down together, and that approximation is worth having, but it is a derivation with a confidence, not a binding. It is marked as such, it carries its own coverage number, and an answer that leans on it says so.

The line If a human at the client could not check the output against something they already know, it is not a transform.