The Six Faces of Bad Data

Six different diseases, six different tests

The vocabulary goes back to Wang and Strong's 1996 study of what data quality means to the people who consume it. Six dimensions survived into nearly every modern data-quality tool, and each one answers a different question about the same cell. The point of this guide is that they are separable: you can break exactly one at a time and watch exactly one meter move.

How this page stays honest: every number is computed in your browser from the 14-row customer table below, plus a small verified contract file for the accuracy check. The audit date is pinned to the constant 2026-07-01, so the timeliness math never shifts under you. Nothing is faked.

Rows marked verified appear in the contract file below, which is what the accuracy meter compares against. Cells outlined in orange are flagged by the rules you build in the next section.

The contract file (ground truth for accuracy)

Accuracy is the only dimension you cannot compute from the table alone. It needs an external source of truth. Here that source is a set of six customers whose name and email were verified against signed onboarding contracts. The accuracy meter compares each verified field in the table against this file and counts mismatches.

Build the rules that catch it

Meters diagnose; rules enforce. A validation rule is one column plus one assertion, evaluated against every row. Build at least 2 rules and watch the offending rows light up in the table above. Note that null values are skipped by every rule except not-null: each rule tests exactly one thing, which is what makes a failing rule diagnostic.

The trap: validity is not accuracy

Inject the "Drift from the truth" defect and look closely: jon.smith@acmecorp.com passes the email regex, is unique, agrees with every other field, and was updated recently. Five meters stay green. Only accuracy falls, and only because this page happens to have a contract file to compare against. That asymmetry is the most expensive lesson in data quality: validity checks are cheap and local, while accuracy needs an external source of truth that usually does not exist for most of your columns. In practice you automate the five cheap dimensions as rules in your pipeline, and you buy accuracy where it matters with audits, double entry, or reconciliation against an authoritative system.

The reverse confusion also costs money: a stale record is not invalid. Every aged last-updated date in this table still parses and sits inside the allowed range, so an in-range rule will never catch staleness. If a meter feels redundant, inject its defect and check which rules fire. Where none do, that dimension needs its own test. For what to do after the diagnosis, the next guide walks the exploratory workflow that turns a suspect table into an understood one.