Broken data announces itself. A missing field, a failed validation, an error log lighting up — someone notices and fixes it.
Almost correct data is different. It passes every check. It loads cleanly. It looks fine. And somewhere downstream, a decision gets made on numbers that are off by a rounding rule, a timezone, a field that meant one thing in the source system and something slightly different in yours.
The system didn’t fail. The process didn’t fail. The gap was smaller than the noise, so no alarm fired.
This is where most data quality problems actually live — not in catastrophic failures but in the quiet drift between what a value represents and what the receiving system assumes it represents. A status code that carries different weight on each side of an integration. A date with an ambiguous format that almost always parses correctly.
Good process design treats “almost” as a category, not a comfort. It asks: what does this field mean, and does both sides agree? It builds checks against business logic, not just syntax. It makes the cost of ambiguity visible before ambiguity becomes a decision.
Scott Andrew Alpaugh puts it simply: if your validation only catches broken, it won’t catch wrong.
The goal isn’t zero errors in the logs. It’s a system where “clean” actually means accurate — not just structurally intact.