Revived: a four-seat dark factory for Pocketful

Created by team Revived on October 06, 2026
Pocketful

Revived is a generic dark factory: four Claude Code seats in one Band Desktop room that take a spec and ship a working service, stage after stage, without a human in the loop. In our submitted Pocketful run the only human input was one task per stage, four messages in total. All four stage folders claim their stage. The run took 2 h 18 min and cost $35.72 at list price. Each seat owns one thing. A Sonnet coordinator hands out work, reviews diffs and rules on disputes. An Opus tester writes a contract and hidden scenarios from the spec alone; no other seat can read them. The Opus builder writes the product and only ever sees the spec sentence it fails, never the test, so it cannot fit the code to the tests. An Opus designer publishes the design system and exact markup before any screen exists, owns the stylesheet and artwork, and checks every screen in Chromium at phone and desktop width. Review changes the work, and the room log shows it: the designer caught a blank cold-load screen and the builder fixed it 79 seconds later; the coordinator's own code review caught an unguarded type assertion that could crash with a 500. We also keep the miss: in stage 4 the coordinator ruled wrong on snapshots surviving an upgrade, and FACTORY.md says so. We measured before we designed. With the shipped tests hidden, cheaper builders did worse and cheap testers gave false confidence, so the builder and tester are frontier models and every rule that mattered became a script: gate.py, coverage.py and layout_check.py. The same mandates, unchanged, ran tablekeeper stage 1, and the first commit passed 120 shipped checks. The repo holds the four stages, the mandates, the tools, FACTORY.md and the full room log.

Category tags: