DarkHarness: Tablekeeper

DarkHarness is a dark-factory harness that runs three AI seats in one Band room: - a coordinator, which splits the work and hands over the complete task and specification; - a builder, which implements and commits each stage; - an independent reviewer, which checks a fresh snapshot of every commit and sends defects back for repair. The submitted run. One human message assigned all four stages. All four were independently accepted within 1 h 58 m, and the coordinator's final report came at 2 h 03 m. The final revision passes all 158 shipped checks. What review added. The reviewer rejected work twice. Both times the revision had passed every shipped check, and both times it was a real defect: - an out-of-range number that broke atomic receipt saving; - a retry that showed the old table after a seating replan. Core design idea: tell dangerous failures from simple ones. - A failure that provably did nothing is retried within fixed limits. - A failure with uncertain effects is never replayed. It is held and reported. - A provider safety refusal is never retried. Getting here took four fresh attempts, and every stop became a fix in the factory. In attempt three, OpenAI's cyber filter refused our reviewer's import-validation re-check. It was a false positive, and we did not try to get around it. Measured cost. About 25.2 M native seat tokens, mostly served from cache. No automatic failure recovery was needed. FACTORY.md in the repository covers the full design, failures, costs and setup. Team Universebaby (UOS). The team wrote the documentation in Korean and translated it into English with an AI assistant.