Cole's dark factory: holdout validator and a perimeter the factory can't amend
Aug 14 coleam00: Archon cron factory — dumb label dispatcher, independent validate-PR holdout, governance files auto-reject; humans file issues and promote releases only.
coleam00/dark-factory-experiment (holdout/ratchet commits Aug 14; live RAG chat app) is a public level-4 dark factory: humans file issues and promote releases; Archon workflows on a cron own triage → implement → validate → merge. Three layers stay separate — Archon (deterministic workflow engine stitching agent nodes to gh/scripts), Claude Code (tools), model (Sonnet/Haiku, swappable). Shared state is GitHub labels, not a private DB: issues march factory:triaging → accepted → in-progress; PRs needs-review → approved (auto-merge) or needs-fix (max two) → needs-human.
The non-negotiables are the thesis receipt. The validator never reads the implementation plan — StrongDM-style holdout against the issue outcome, so the maker cannot game acceptance. Governance files (MISSION.md, FACTORY_RULES.md, CLAUDE.md) plus auth/rate-limit/deploy configs are a human-authored perimeter: any PR touching them auto-rejects before other checks. The dispatcher is dumb on purpose — ~100-line bash every 30 minutes, priority fix-PR → validate-PR → implement → triage, MAX_PARALLEL=4; an earlier LLM dispatcher hallucinated work that did not exist. Finishing in-flight work before new triage is load-bearing. Weekly browser regression files issues back into the same queue. Flood protection and per-node maxBudgetUsd cap spend. Explicitly not level 5: the factory does not invent its own product backlog.
Why it matters for code review
Overnight harnessed momentum compounds only when merge authority is a holdout against reality and the factory cannot rewrite the rules that judge it. Continuous factories beat burst Claude heroics when the cron keeps verified PRs moving — and when "done" is an independent validator, not the implementer's story about its plan.