About

Lights On, Lights Off is a textbook about AI software factories as they stood in July 2026 — what a loop, a harness and a factory actually are, what “dark” claims, and why verification rather than generation is the thing that decides all of it.

It exists because the material was sitting in two long research documents and a scatter of design notes and retro reports. Nobody learns from those.

Written by an agent pipeline

This site was written by an AI agent pipeline, and reviewed by machine judges before a human read it. That is stated here, in the colophon, and on the case-study pages, because it is both the honest disclosure and the point: the site is an instance of the arrangement it documents.

The stages it went through are the ones described in Five factories in one workspace:

1. Brief — purpose, two reader personas with explicit bounce conditions, archetype, content model, provenance rules, and the page inventory that defines done. Human-approved.
2. Design — a named design system argued from the content, built as a static mock and approved before a single page was scaffolded. Rejecting a mock costs minutes; rejecting a built site costs the whole loop.
3. Scaffold — the harness: deterministic floor, screenshot capture, the register validator.
4. Build loop — the pages written, then judged each round by two independent agents, one reading the rendered screenshots and one reading the prose. Capped rounds, then a human.
5. Publish — human-gated.

The human gates were the brief, the design and the final readworkspace. Everything in between was decided by machines, which is roughly a Level 3 to 4 arrangement on the ladder — the same rung as the organisations in the field report, on a smaller and much lower-stakes object.

What is guaranteed, and what is not

Every number, date, percentage, dollar amount, organisation name and attributed coinage in the chapters carries an evidence chip pointing at an entry in the register. A validator runs as part of the build and fails it if a claim points at nothing, so that guarantee holds by construction rather than by care.

What no machine here checks is whether a source is fairly characterised. It is the gate this site cannot automate, for the reason chapter 8 gives: it is the check that can be faked out by the same process being checked. The register page says so on the page rather than in a footnote.

Three further limits, stated plainly:

• The field survey is not complete, and does not claim to be. It reports what the sources record.
• Workspace evidence is not independently verifiable. The case studies draw on repositories on the author’s own machine. They are cited by path and read date, they carry their own tier, and the pages that use them say what they cannot prove.
• The sample in the darkening case study is three runs, one of which reached a human review. Counts are stated with that number attached.

Sources

Two research documents, both in the repository under specs/: a survey of the field as of mid-2026, and a conceptual frame after Addy Osmani’s Software Factories, Light and Dark and Dex Horthy’s Harness Engineering is not Enough, with an original darkening analysis. The case studies additionally draw on five pipeline repositories and their retro reports.

Colophon

Design system: Control Room — the plant floor seen from the room the human moved to. The light theme is the lit factory; the dark theme is the lights-out floor. Amber marks a gate a human holds and steel marks a gate a machine holds, on every diagram and every ledger, so a figure can be read without a legend.

Set in Archivo (signage grotesque), Newsreader (reading) and JetBrains Mono (every machine record). Figures are inline SVG keyed to the palette tokens, which is why they re-light with the theme. No generated artwork: a generator that invents text on drawn signage would be a fabrication surface on a site whose argument is about verifiable claims.

Built with Jekyll, published on GitHub Pages. Written by Claude (Opus 5) in an agent pipeline, directed by Jörn Dinkla.

Proposing an update

The updates page sets out what gets recorded, what does not, and the mechanics that keep the log honest. Corrections to a source’s tier are the cheapest and most valuable contribution: a tier is defined in one place and every chip that cites it follows.