Five sites, one site
The site you are reading is the survivor of a five-way benchmark. In July 2026 the same brief — this book — was given to five model/effort pairs, each of which built a complete site. The five were compared, one was chosen as the base, the strongest material from the other four was grafted onto it, and the four superseded sites were taken down. This page is the record of that process, and deliberately so: once the variants went dark, it became the only citable evidence the benchmark happenedworkspace.
The setup
One brief. Five sessions, differing only in model and reasoning effort. No variant saw another’s output while buildingworkspace.
| Variant | Model / effort | Verdict in the mergeworkspace |
|---|---|---|
| baseline | Fable | Quarry for visual assets only; its clean minimalism was already subsumed by the base. |
| alt | Opus 5 · high | Donor: the staged playbook with stop conditions, the staffed/unread figure pair, the provenance footer, the tier-license prose. |
| codex | GPT-5.6 · medium | Donor: the dark-after-design finding stated as a claim, chapter role labels, the position pull quote. |
| field guide | GPT-5.6 · high | Donor: the pedagogy layer. The 90-second argument, the dimmer, the calibration status labels, the large social card. |
| om | Opus 5 · medium | Base. Architecture, the evidence register with its build gate, the durable/dated split, the reading ladder, the chapter spine. |
What the comparison found
Structure decided it. The base won on what the merge spec calls structural integrity: an evidence register the build enforces, a URL structure that separates what is durable from what will age, and complete canonical metadata, plus the conceptual spine of the reading ladder. Everything else could be grafted on afterwards; a site’s architecture cannot beworkspace.
Effort did not buy structure. The comparison’s oddest row: on this brief, the Opus 5 medium run beat the Opus 5 high run on structural integrity — the extra reasoning effort went into richer side registries and a more elaborate page apparatus, which the merge then dropped as over-elaboratedworkspace. Meanwhile the GPT-5.6 high run produced the best teaching layer on the thinnest structure, and the GPT-5.6 medium run produced the single sharpest empirical sentence in the corpus, the dark-after-design finding, inside the weakest shell of the fiveworkspace. Model and effort moved quality along different axes for different models. One brief and one run per pair is a sample of one, five times over; read it as an observation, not a curve.
The graft list is visible on this site. The 90-second argument and the position quote on the front page; the dimmer; the playbook’s stop conditions; the staffed/unread pair in chapter 3; the status chips on the five-factories table; the role label in every chapter’s kicker; the provenance line in the footer of every page. Each was argued from a specific donor in the spec, with the conflicts decided in writing before the merge session startedworkspace.
What happened to the evidence
The four superseded sites were unpublished rather than left up, because five near-identical sites answering the same brief under one author is reader-hostile. GitHub Pages does not redirect an unpublished project site, so their URLs simply stopped servingworkspace. Before that, the built output of all five and home-page captures of the principals went into the repository’s archive, which is what the chips on this page point atworkspace.
This site then took over the canonical URL. Inbound links to the old baseline’s chapter slugs land on redirect stubs pointing at the corresponding merged chapter; links into the other three variants were accepted as losses; they were short-lived benchmarksworkspace.
The recursive point
The other two case studies describe pipelines this workspace built. This one describes the site itself being put through the pattern the book teaches: several makers, an explicit comparison, judgment applied at the level where it is cheapest (choosing between finished structures rather than steering any single build), and the losing work archived rather than silently overwrittenworkspace. Chapter 8 calls the human’s job designing the line and guarding the gate. Merging five sites into one is that job, applied to the book about it.
It demonstrates judgment-as-selection: five cheap parallel attempts and one comparison turned out to be a better use of the same attention than five rounds of feedback on one attempt. What it cannot test is anything about models in general — five sites, one brief, one comparing reader, and the comparison performed by the workspace that built every candidate. The effort-response observation in particular is one data point per pair, and this page says so each time it comes up.