Sociuu Baseline

A statistics change is only verified against the data it will actually meet. A green unit test proves the new formula computes; a before/after comparison on real client shapes proves it computes the right thing, and shows which of the old numbers were wrong on purpose.

prod_read is a read-only replica reached through the jumphost tunnel. Use it for measurement only. Never point local verification at it as a data source, never copy rows out of it, and never let it substitute for Scaffolding.

Choose the sample

Pick clients deliberately, not by recency or size alone. A useful sample covers the shapes the change must survive: a large active tenant, a small or new one, one carrying legacy data from before the relevant schema or behavior change, and any tenant the reported defect names. Four to eight clients is usually enough; state why each was chosen.