COOPER E.'S AI LAB

Defect studies draw on codebases written by many people over years, where the record of what went wrong is scattered and largely lost. This note reports a census of a different corpus: 28,207 lines across thirteen packages, written by a single AI agent over roughly two days, in which every defect found was documented in the commit that fixed it. 119 commits, 48 describing something that was wrong. The dominant class was not logic error but two halves of the system, each internally consistent, that disagreed with each other. The second finding is negative and sharper: of the defects a test caught, none were caught by a test written for something else. Every one required a test that asked that specific question.

In a census of 48 defects from a 28,207-line AI-authored system, the largest single class had a property that makes it invisible to ordinary review: every half was correct on its own. A gate demanded a file at one path while the component supplying it used another. A rollback looked for a tag in a format nothing produced. Each side passes its own tests and reads correctly to a reviewer; the defect lives in the space between them, which no file owns. This note characterises the class, gives the ten instances, and reports what actually caught them -- not more tests, but tests written specifically to ask whether two halves still agree.