Measuring the sloppiness of code

Code can pass tests while becoming harder to extend. The essay on measuring code sloppiness explores ways to detect that deterioration, including the share of duplicated or flagged lines and the share of complexity concentrated in a few high-complexity functions. These measurements make aspects of maintainability visible that a functional pass rate can miss.

The underlying SlopCodeBench research evaluates agents repeatedly extending their own previous solutions as requirements change. That setup makes early design choices matter to later work. Its measurements track how code structure evolves across those changes, giving developers a way to investigate accumulating maintenance problems rather than judging each generated feature in isolation.