🤓 Latest Submissions
Twin Builder Never Grades Itself
The agent that builds a feature is usually the one that decides it is done. Twin is a dark factory in BAND built on one rule: the builder never grades its own work. Six seats on three model families. The coordinator (Claude Code) splits the work and never writes code. The builder and surface seats (Claude Code) write the service and the screens. The modeler (Codex) writes an executable model of the spec without ever reading the product. The gatekeeper (Codex) runs the same operations against product and model, fires concurrent requests, plants faults to prove its own checks catch them, and is the only seat that can ACCEPT. The auditor (DeepSeek on Featherless) reads the spec a third way. The judged run: one dispatch, 0 human messages after it, all four Pocketful stages accepted in 3 h 19 m. The gatekeeper rejected work 12 times, and every rejected stage was later accepted on a newer writer revision. One rejection came while the organizers' checker passed the same revision 147 of 147: the export timed out at 10.01 s; the fix landed 137 s later and the export answered in 0.011 s. Proof you can check: the organizers' harness on a fresh clone passes 147/182/188/193 provided checks; a sealed holdout committed before the run scores 73/73; the modeler's ledger quotes 513 spec sentences; every commit to a stage folder is by a seat, and the gatekeeper accepted stage 2 with 82 screenshots at 375 and 1440 px. Cost: Band's list-price estimate $204.51, Featherless $1.03. Start with JUDGE-GUIDE.md in the repo; the Factory Floor replays the whole room.
6 Oct 2026
