The Dark Factory: Cheapest Hands First

The Dark Factory is a three-seat agent team in Band Desktop that turns a written spec into a staged, containerised service from one human message. We pointed it at tablekeeper, a restaurant booking system that must never double-book a table under concurrency, retries and time zones, and left the room. Two ideas drive it. First, the shipped checks are a sample; the spec is the test. Only 11% of stage 3 tests ship, so our architect (Opus 5.5) turns every spec sentence into a numbered requirement and writes a black-box test suite before any code exists: wrong JSON types, stray whitespace, exact fields, retries, bursts of 50 requests, daylight-saving edges, data upgrades, every screen at 375 px. Second, the cheapest model that passes wins. Each task climbs a ladder: Haiku 4.5 gets two attempts, Sonnet 5.5 two more, and the architect finishes only if both fail. The architect re-runs every claimed success itself, so a cheap model can be wrong, just never wrong and accepted. In the submitted run Haiku finished 4 of 9 tasks, all first try, Sonnet 5, Opus 0. The room shows real teamwork: the architect caught a Haiku UI at 370/390 and sent a product review of its screenshots; Sonnet rewrote it to 391/391. Later the fixer proved two of the architect's own tests contradicted the spec, and the architect agreed and fixed them. Results: all 4 stages pass the official checks in isolated Docker, 265/266 of our spec-derived tests, 5 h 26 min, $27.46, about half the cost of our earlier five-seat factory at the same quality. Mandates never name a product or endpoint. FACTORY.md, mandates, room.json and 16 measured runs are in the repo.