How do you evaluate synthetic data when there's no blanket metric?

Open question. Most quality discussions collapse into "is the marginal distribution close enough to real" — which says little about whether the data will hold up the use it's meant to support. A synthetic vendor-master that's perfect for testing a deduplication pipeline can be useless for testing a fraud detector, and vice versa. I don't yet have a clean formulation of what the bar should be when the answer is always "depends on the downstream task."