Tablekeeper Dark Factory: tests that get tested

In a dark factory nobody reads the code, so the tests are the only reviewer. We built a factory that measures its own tests. Four Codex seats share one Band room. The lead splits the spec and closes each stage. The examiner writes a hidden exam from the spec: API tests, browser journeys on phone and desktop, and model-judged look checks. The builder writes the code and never sees the exam, only failure reports that quote the spec. The auditor plants bugs from the spec into copies of each accepted build and counts how many the exam catches, next to how many the hackathon's public checks catch. When a bug survives, the examiner gets its spec section and sentence, never the bug, and writes new tests. Judged run on Tablekeeper: one human message started it, and the seats finished stages 1 to 3 on their own. Across those stages we planted 28 bugs. The public checks caught 12. Our exam caught 27 on the first pass and all 28 after one follow-up. The builder's first commit passed the exam in every stage. Band estimated the room at 19.8 million tokens and $2.45, on a ChatGPT plan with no API billing. Stage 4 was still in progress at the deadline. The repo holds the code, the Band room export, and a scorecard for every stage that links each number to its source file.