chapter 06 · first-hand
Case study: our factories
This workshop runs four factories: one for software products (spec to tested implementation), one for interactive demos, one for visual novels, and one for websites, the last of which produced the page you are reading. None of them ships code to production systems; all of them ship artifacts under a real name to a public URL. They are small enough to hold in one head, which makes them useful as a laboratory: every claim in this chapter is our experience, checkable against the run journals in the repositories.
The website factory drives an idea from an empty directory to a published site through five gated stages: brief, design (a static mock, approved before any scaffolding), scaffold, a build/judge loop, publish. Its verification model is the three-way split from chapter 3: a deterministic floor, two independent judges (vision over screenshots, editorial over prose), and a human feel check. One detail matters more than it looks: the editorial judge runs on a stronger model family than the maker. Prose tells track the writing model’s own distribution, so a judge from the same family reads the maker’s tells as normal prose and lets them through. Judge-equals-maker fails silently, and it fails precisely on the taste dimensions the judge exists to check.
§ 6.1The darkening exercise
In July 2026 we examined what it would take to run this factory dark after brief approval: no design gate, no feel check, no publish confirmation. The mechanical work turned out to be roughly a weekend: a driver that chains the stages, a third judge for the design mock, a publish policy file, a hard rule that judge-REVISE at the round cap parks the run instead of shipping it. Dark factories need defined terminal states, not questions.
Three things resisted, and they generalize:
No machine oracle exists for factual integrity. Whether citations resolve is checkable; whether they are honest is not. The same generative habit that put invented author names on drawn book spines in one of our image renders (the retro dividend below) will, given the chance, put plausible-looking sources into prose — and this is the one gate where the check can be faked out by the very distribution it is checking. The workable compromise is per-archetype darkness: fiction and public-domain reading editions have almost no fabrication surface; essay sites with factual claims, like this one, do, and stay lit.
Feel is, by construction, whatever the judges cannot see. The gate split defines the human check as their residue. Run evidence says the residue is real: one recent site needed a human feedback round after the editorial judge passed. Deleting the gate does not make the residue zero; it makes it invisible.
Publishing happens under a real name, facing outward. A public URL is reputational and cached once crawled. The cheap compromise keeping ~95% of the darkness: end the dark run at a private repository with the site fully built, judged, and deployed, leaving a single bit, the visibility flip, to a person.
§ 6.2The retro loop is the dimmer
The factory’s outer loop is a retro skill that mines run journals (commit histories) for what humans caught after the judges passed. That is exactly the measurement a dark factory lacks: the human’s marginal catch rate per gate. When retros show the feedback round catching nothing for N consecutive runs on an archetype, that gate has earned lights-out there, with evidence. The same loop shows which gate stays lit forever: the generated sites are disposable, but the factory’s knowledge docs and templates are its brownfield, where comprehension debt would actually accumulate.
§ 6.3Where our switches sit
The standing recommendation, one line: build dark-after-design, terminate at a live private URL, keep fabrication-prone archetypes lit, and let the retro loop’s marginal-catch-rate data decide which gates go dark next. The design gate stays lit not out of caution but arithmetic: a rejected mock costs minutes, a rejected site costs the whole loop, and the lit-factory doctrine of chapter 4 says to spend judgment upstream where it is cheapest. Darkness is earned per gate, per archetype, with evidence. That is the entire method, and nothing in chapters 1–5 suggests it stops being true at larger scale.