The production MLInferenceDeps bundle the scorer runs with — the
RunViewMLModelLoader (loads the MJ: ML Models row), the
MJSidecarPredictor (runs /predict against the Python sidecar) and the
StorageProviderArtifactLoader (reads the model bytes through the storage provider
their MJ: Files row names, the read-side inverse of StorageProviderArtifactStore; it takes
the provider per call, so the bundle needs none). Shared by BuildScoreRecordSetRunner
(the Score action / Remote Op path) AND the startup work-type registration so a model
trained one way is always scored the same way — one source of truth (CLAUDE.md DRY).
The production MLInferenceDeps bundle the scorer runs with — the RunViewMLModelLoader (loads the
MJ: ML Modelsrow), the MJSidecarPredictor (runs/predictagainst the Python sidecar) and the StorageProviderArtifactLoader (reads the model bytes through the storage provider theirMJ: Filesrow names, the read-side inverse ofStorageProviderArtifactStore; it takes the provider per call, so the bundle needs none). Shared by BuildScoreRecordSetRunner (the Score action / Remote Op path) AND the startup work-type registration so a model trained one way is always scored the same way — one source of truth (CLAUDE.md DRY).