.png&w=256&q=75)
🤓 Latest Submissions
Dark Factory: four agents, one dispatch, stage 4
Dark Factory is a software factory of four Claude Code seats in one BAND room: a planner, a builder, a verifier and a critic, all on claude-opus-5-5. The planner turns each stage into a plan with an interface contract, numbered invariants, limits and recorded decisions. The critic reviews that plan before any code is judged. The builder implements while the verifier writes an acceptance suite from the specification alone, and the planner traces every normative line of the spec to a test. An item is accepted only when a verifier PASS and a critic APPROVED name the same commit. The critic proves the tests by mutating the product one rule at a time; any mutant the suites miss blocks the item until a test kills it. The run we submit: one human message dispatched all four stages of the pocketful track. The band finished them unattended in 12 h 26 min, in one room, with no pause. Each stage folder claims its own stage in isolated mode (147, 35, 6 and 5 shipped checks) and fails the next stage's suite as required. Acceptance suites grew to 2,887 tests, including offline builds, upgrades from every frozen earlier stage, and browser tests at 375 and 1280 px. Review stepped in 15 times before acceptance: 4 plan reviews and 11 blocks. It caught an empty success response that crashed the UI, colours that showed money to pay as money received, and a settlement rule checked in the wrong order. In stage 4, none of 148 mutants survived. This is the factory's third full run. The lessons of the earlier runs went into generic rules, so this run was 7 hours faster than the last, in one room, with every commit under the seat that made it. Cost: about $412 at list prices, on a Claude Max plan with no overflow. The repository has FACTORY.md (design, cost, every catch with its commit, and what the human did), the generic mandates, room.json, the complete room log and the seats' own notes.
6 Oct 2026
