Pocketful: an evidence-gated dark factory

Five coding-agent seats in one Band Desktop room. The Coordinator (Claude Code) turns the spec into a requirements ledger: all 273 stage-1 sentences are accounted for (252 obligations, 21 waivers). Builder (Codex) and Builder-Two (Claude Code) build on separate branches and never see the shipped tests. The Checker writes its own checks from the spec: 397 at stage 3's acceptance, against 188 shipped. The design crosses model families; stage 3 was accepted while the Adversary's W-10 review was still pending, and FACTORY.md records that exception. No seat can say "done". The Checker runs tools/gate.py on an exact commit: credential and test-fitting scans, the event harness in isolated mode for every earlier suite, then its own checks. Each receipt is chained by sha256, and only a green one lets a stage onto main. One human dispatch; no human room posts afterward. Outside-room infrastructure repairs are disclosed in FACTORY.md. Stages 1-3 were accepted; a fresh isolated clone passes 147/147, 182/182 and 188/188. 559 of 572 obligations (98%) have a Checker-written check: referenced coverage, not proof of correctness. The catch in the video: RED #8, a defect the shipped tests passed, then GREEN #9 14 min 44 s later on a candidate that was never merged, so submitted stage 3 still has that disclosed defect. What failed: vendor safety refusals stopped the red-team seat, the factory had no liveness check, and stage 4 was built but never accepted.