Base64 of the serialized artifact, ready for the sidecar PredictRequest.
Point-in-time strategy frozen on the model's pipeline (assembly input).
The model's frozen, ordered feature schema (the inference input contract).
The frozen FeatureStep DAG (assembly input).
Frozen fitted-preprocessing params — passed through to /predict unchanged (§6.2).
Leakage guard frozen on the model's pipeline (assembly input).
The model primary-key id (cache key + lineage).
What the model predicts (classification / regression).
Ordered source bindings the model's features draw from (assembly input).
Target entity whose records this model scores (training-unit entity).
The label/target column name (diagnostics + lineage; never a feature).
The frozen inference contract resolved off a loaded
MJ: ML Modelsrow — the artifact bytes plus everything the sidecar needs to apply (never re-fit) the model. Cached warm per processor run and reused across every record in a batch.