Data Quality Audit & Reporting Tool

Cleaning scripts are trusted because their output looks tidy: no blanks, no duplicates, no unparsed dates, job done.

But a wrongly repaired cell is well formed, in range and consistent with its neighbours. You cannot measure a cleaner's error rate by looking at its output.

A synthetic file with known defects is generated: 2,400 people, 14 columns, 15 kinds of defect injected into 22% of rows. The same dirty file goes to two cleaners, the rule engine here and a competent, ordinary twenty-line pandas cleaner. Both are scored against ground truth.

The whole architecture follows one line: every rule returns a confidence, and any suggestion below 0.85 is not applied. The row goes to a human instead.

Cell accuracy is 0.9958 against 0.9893, but the number that matters is different: 10 silently wrong cells against 361. The naive cleaner's output looks cleaner (no blanks, no review queue) while passing 36 times as many errors in silence.

Whether the decision changes was measured too. A 4.1% pay gap was built into the population by construction: the clean population reads 3.85%, the rule engine 3.86%, the naive cleaner 3.37%. The naive cleaner understates the cost of closing that gap by 366 thousand units.

The measurement found three defects in the tool itself, none of them visible from its output: 61 correct salaries multiplied by twelve, 28 required fields silently blank, 8 malformed email addresses with no log entry at all. 134 tests.