Tablekeeper Dark Factory — three AI seats

What it is. A reusable "dark factory": three AI seats in one BAND room turn a written specification into a running, tested service without a human in the loop. The human posts exactly one task message per stage; everything after that is the band. The seats. Architect (GLM-5.3) turns the specification into a numbered requirements ledger, writes work orders, and reports after acceptance; it never writes code. Coder (MiniMax-M3) implements the service, Dockerfile, run instructions and tests. Reviewer (DeepSeek-V4.1-Flash) designs probes in parallel, then verifies a fresh clone in a clean offline container, runs the official check, and accepts or rejects with a reproduction. Why unattended. A reply gate stops acknowledgement ping-pong (our first rehearsal burned ~$156) and gives an "announces and stops" seat a bounded nudge. Seat shims refuse human-identity CLI posts; one gap was found in the submitted run, documented and fixed. An owner conductor posts each stage task and verifies the official isolated check before the next; it never nudges. An independent guard pulls seats out if message rate or spend spikes. Results. Official harness: claimed stage 3 (suites 1-3 pass: 120/120, 25/25, 7/7; suite 4 4/6), with every stage folder healthy in a clean container without network. Reviewer accepted stages 1-2; stage 3 was committed but not reviewed before context compaction ended its turn. The Reviewer rejected 3 handoffs from its probes and browser checks. Total time was 4.5 hours (05:28-09:55 WIB), with $26.64 model spend on Nebius Token Factory. Reusable. Mandates are generic: track detail travels in each stage task, so the same factory can target another problem by changing only that message.