Four factories on one machine

A lit factory small enough to describe completely, including the parts that failed. It is the only evidence on this site that was not read out of somebody else's write-up.

Everything else here is compiled from published accounts. This page is not. It reports a four-plugin factory suite built and operated by the author of this site, included for one reason: published factory write-ups are almost always success stories, and success stories leave out the gate that had to be repaired.

The obligation that comes with that is to be equally clear about what this evidence cannot support, which is the last section.

The four

Each is a Claude Code plugin that drives an empty directory to a finished artefact through gated stages. They share two invariants deliberately: worker agents never hard-code build tools — they read a Commands table in the generated repository — and no agent reviews its own output.

Factory	Product	What the checker can be
first-factory	software	a green gate: BDD suite, per-task reviewer, QA agent
demo-factory	visual demos	a deterministic floor plus a vision judge over screenshots
visual-novel-factory	playable novels	hard gates on the story graph, an editorial judge on the prose
website-factory	published sites	floor, two independent judges, then a person

That third column is the whole design. The first factory rests on a machine-decidable definition of done. The other three do not have one — “does the layout breathe, is the sequence comprehensible” is not a Cucumber assertion — so they were given a different arrangement rather than a weakened version of the same one, because a green gate over editorial work is ceremony.

The pipeline, and where its lights are

1 Brief · humanA person approves the purpose and the page inventory that defines done.
2 Design · humanA person approves a static mock before anything is scaffolded.
3 Scaffold · machineFloor green: build, links, screenshots.
4 Build loop · machine, then humanTwo judges, at most three rounds, then a human feel-check closes it.
5 Publish · humanA person confirms the name, the visibility, and that it goes live.

human gate machine gate machine, then human — the capped box

Fig. 8.1Three of the five gates are human, and the two that are not still hand back to a person before anything ships.

Three of five gates are human, and they are not the three you would pick to save time. They are the three where a wrong answer is expensive and only a person can catch it: what the site is for, what it should look like, and whether it goes out under a real name.

What the runs show

10 sites published through the full gated arc, 2026-07-11 to 07-23 as of 2026-07-23 First-hand 2 of 3 judge rounds used on the most recent run before both judges passed as of 2026-07-23 First-hand 1 human feedback round that still caught something after both judges had passed as of 2026-07-23 First-hand

The verdict pattern on the most recent run is the healthy one: no round where both judges passed immediately — an all-pass first round usually means the rubric is too loose rather than that the work was perfect — and the judges caught two defects the maker had missed. One of those is worth describing, because it is a small lesson about what a judge is for.

The bug that only the pixels showed

Illustration plates rendered in the wrong aspect ratio, clipping the compositions. Inspecting the page's computed styles reported the aspect ratio as correct, because it was correct in the CSS — an explicit height attribute on the image element was silently winning instead. The defect existed only in rendered pixels. A DOM-reading check would have passed it forever; the vision judge caught it in one round.

The correction went further than the site. The factory's own craft documentation had been teaching the failing pattern, so the fix was applied upstream to the knowledge document and to both stack templates. That is the outer loop working: the run improved the factory, not just the artefact.

The gate that had to be repaired

The editorial judge is pinned to a different, stronger model family than the one drafting the prose. That pin is a repair, not a precaution. A judge drawn from the maker's own distribution approved prose it should have flagged — the tells were unremarkable to it, because it produces them too. Every individual verdict looked considered while the gate did nothing at all.

This is the failure mode from Station 06 in its natural habitat: a check that is only as independent as its failure modes, failing silently on the one dimension it was built for.

The outer loop is the dimmer switch

The most useful component turned out to be the least glamorous: a retro skill that sweeps run histories, clusters fixes that recur across runs, and proposes edits to the rubrics, templates and knowledge documents. The 2026-07-23 report proposed six changes; all six were accepted. Three of them corrected documentation that had been teaching a failing pattern.

That loop also produces the one measurement a dark factory has no way to obtain: the human's marginal catch rate per gate — what a person still finds after the machines have signed off. When that number has been zero for several consecutive runs on a kind of work, that gate has earned lights-out there, with evidence. Until then, removing it does not reduce the risk; it removes the instrument that was measuring it.

What this evidence cannot support

Stated plainly, because a case study written by the person who built the thing has an obvious defect.

One operator, one context. There is no team, no on-call rotation, no second opinion that was not requested. Nothing here generalises to coordination problems.
The output is disposable. These are content sites. Nobody is paged when one is wrong, and the cost of a bad run is a rerun. Almost every hard question about darkness concerns systems where that is not true.
The numbers are small. Ten runs is enough to notice a pattern and nowhere near enough to measure a rate. The judge statistics above come from a single run's journal.
The repositories are unpublished, so every claim here is cited by file path rather than by link, and you cannot check it. Treat it accordingly — it is First-hand, which is a statement about proximity, not about verifiability.

What it does support is narrower and, I think, still worth the page: that the gate split is operable by one person, that judges catch things makers miss, that a judge sharing the maker's distribution can fail silently — and that the human gate at the end was still earning its place on the most recent run.