Updates

This site is a snapshot with a maintenance plan. The chapters are written to survive; the dated snapshot pages are not, and this log is where the difference gets recorded.

1. 2026-10-05 correction amends: case studies, updates

   Predecessor drafts taken offline

   The second-edition entry below said the four superseded drafts were unpublished. That did not happen until today: they stayed live for ten weeks after the merge. They are now offline and their repositories archived.

   Removed: the Five sites, one site case study. It compared the models and reasoning-effort settings that built the drafts. Those models are no longer current, and a comparison of retired models teaches little about factories. The archive and the merge spec remain in the repository as the recordworkspace.

2. 2026-07-26 publication amends: all pages

   Second edition: five sites merged into one

   This site became the canonical ai-factories site. It began as one of five variants built from the same brief by different model/effort pairs as a benchmark; the comparison chose this one as the base and grafted the strongest material from the other four onto it, per a merge spec written before the session that executed itworkspace.

   What arrived in the merge. The 90-second argument and the position quote on the front page; the dimmer in chapter 8 and standalone; the playbook with per-stage stop conditions; the staffed/unread figure pair in chapter 3; calibration status chips and per-factory lesson lines on the case studies; the dark-after-design finding stated as a claim; chapter role labels; the provenance footer on every page; a large social card, with a build gate that fails any page missing its metadata.

   What happened to the predecessors. The four superseded sites were unpublished. GitHub Pages does not redirect an unpublished project site, so their URLs stopped serving; the old baseline’s chapter slugs redirect here to the corresponding chapters. The built output of all five variants was archived into this repository firstworkspace.

   Recorded, not overwritten. The first edition below stands as published; this entry is the reversal-safe way this log records a change of that size.

3. 2026-07-26 publication amends: all pages

   First edition

   The site goes up: eight chapters, two dated snapshots, two case studies and the apparatus — register, glossary, this log.

   Where the material comes from. Two research documents assembled in July 2026: a survey of the field as of mid-2026, and a conceptual frame after Osmani and Horthy with an original darkening analysis attached. Both are in specs/ in the repository. The case studies draw on five pipeline repositories on the author’s machine and their retro reports, cited as workspace sources.

   What is deliberately absent. No generated artwork — the figures are inline SVG keyed to the palette. A generator that invents text on drawn signage is a fabrication surface, and a site arguing for verifiable claims should not carry one.

   What will date fastest, in order: the toolbox, then the field report, then chapter 6’s benchmark section. The concept chapters should survive; if they do not, that is the interesting failure and it belongs in this log.

How this site is kept current

The mechanics are deliberately boring, because the failure mode of a living document is that nobody can tell what changed.

• Source ids are immutable. An entry in the register keeps its id forever. Corrections edit the entry's fields or add a new entry; ids are never reused, because they are public anchors.
• This log is append-only. An update names what it amends and does not rewrite history. If a chapter's claim turns out to be wrong, the correction is a new entry that says so, not a silent edit.
• The dated pages carry a snapshot banner driven by one date in the site configuration, so the stamp cannot drift out of sync with itself.
• Every chapter carries a reviewed date in its rail. A chapter whose date is old has not been checked recently, which is a different statement from "is wrong" and a more useful one.
• The build enforces the citations. Adding a claim without a registered source fails the build, so the register cannot quietly fall behind the prose.

What would justify an update

Not every development is a change to this site. The bar, roughly:

• A snapshot page's table becomes false — an organisation moves level, a volume figure is superseded, a tool named here is discontinued.
• A contested source is resolved in either direction, or a vendor-reported figure gets independent measurement. Both change a tier, which is a one-line edit that propagates everywhere.
• Someone publishes a marginal catch rate — what a human gate still catches after the machines pass. Chapter 6 argues that is the measurement the field is missing; if it arrives, that chapter changes.
• A concept chapter's argument is falsified. That is the update worth writing.

A new benchmark score is not, on its own, a reason to touch this site.