How a Fake Default Number Silently Corrupted Three Production Systems

A software engineer discovered that a hardcoded default value of 1.9 percent — originally added to prevent a dashboard crash — was quietly distorting results across three separate production systems, including an ad-experiment engine, a Medicare billing tool, and a denial classifier. Because the number appeared plausible, it survived two years of code reviews without being flagged. The flawed baseline fed sample-size calculations that could be off by a factor of 20, causing experiments to either end too early or run far longer than necessary. The author's fix was not a better estimate but a structural refusal: systems now return errors or route to named human-review states when ground truth is unavailable. Across the billing and denial systems, replacing missing categories with explicit schema states lifted classification accuracy from 83.3 percent to 90 percent and mapped-action accuracy from 36.7 percent to 76.7 percent.