The kept levels of each categorical column, by column name; scoring folds every other value into the fold level.
The typed columns the matrix was assembled from, in order.
That snapshot's content hash, so the lineage names the exact rows even if the row were ever re-pointed.
The MJ: ML Training Snapshots row the model was trained from.
The entity whose records the unit rows are (the entity a scoring scope must hold).
The unit version a model was trained on, the typed feature columns its matrix was assembled from, the kept levels of each categorical column and the Training Snapshot its rows came from.