Golden Record

Four systems hold a customer. The addresses differ, two have different spellings of the name, one has a phone number the others lack, and none is wrong — each was correct in its own context at the time.

A golden record is the reconciled version. Producing it needs two things and the second is where the work is. Matching decides which records refer to the same real entity, which is probabilistic in the absence of a shared identifier and has to be tuned: too loose merges two genuine customers, too tight leaves duplicates. Survivorship decides, field by field, which source wins — most recently updated, most trusted source for that field, most complete, or a human decision.

Two design decisions determine whether it is usable. Whether the golden record is pushed back to the source systems, which is architecturally cleaner and politically hard, or held centrally and referenced, which is easier and leaves the sources inconsistent. And whether merges are reversible, because they will sometimes be wrong and unpicking a merge without a stored history of what was merged is close to impossible.

The failure mode to plan for is that the match confidence threshold is a business decision with real consequences — a false merge in a healthcare or banking context is a serious incident — and it is usually set by whoever configured the tool.