Every number we publish, and how we got it.

When the visitor clicks “Open the … site” on the trade page for their business, for example /for/cafe.

When the warm-up screen has cleared, “Get this site” is ready, and the preview on screen shows the business name (and, in the edit run, the typed phone, email and paragraph) with the cover photo at the top set to the photo chosen in the editor, and that photo loaded and decoded. No run went further: nothing was bought and no checkout was opened.

The trade link opens the builder straight into setup: two setup steps (business name, starting look) left on their suggestions, then Build my website and the wait for the warm-up screen. On a phone, setup ends with the site already on screen.

The same, plus a business name typed in setup, then a phone number, a contact email and one paragraph (the About story, about 170 characters) typed in the editor, and the cover photo swapped for another photo from the free library.

1.5 s to decide before every click or tap, and typing at 5 characters a second. Reading time beyond that is not modelled.

The “automated” column is the same steps without simulated reading or typing time. It is not pure builder time: it includes the test harness's own waits (letting a panel stop scrolling, closing a pop-up, re-checking a tap), measured in every run at up to 2.9 s (median 1.0 s).

desktop 1440×900; phone 390×844 (emulated, touch, DPR 3). Headless Chromium (Chrome/153.0.8010.12).

cafe, plumber, salon, accountant, gym; Australia and United States. A fresh browser with no saved draft for every run.

The same internet connection for every run (the tester's own line, not a data centre), with no throttling. Measured page loads: the trade page in 0.8 s (median), the builder in 2.4 s (median; 1.1 s to 32.8 s), with 853 KB transferred for the builder page itself.

240 timed runs, 240 completed (8 needed a second attempt because the first attempt stopped with an error).

In 62 runs the editor's “find your voice” invite slid up mid-edit and had to be closed, sometimes taking a tap meant for the photo; those closes and re-taps are counted in the time, as they would be for a person.

Up to 4 ran at once on one computer. Analytics and advertising measurement was switched off so the runs were not counted as visitors.

How long a real person takes: nobody was timed, and that measurement is still not made. It also stops before checkout, so it says nothing about paying, the domain, or the site going live.

A slower connection or an older phone will be slower, mostly in the page load and the warm-up screen, which waits for the photos and runs between five and fifteen seconds by design.

Revalidation required. Measured against the builder eaa6cd33fef5 (site commit c81bc4c28); the builder is now 59d306177370.

Method and raw timings: tools/astra/probe-speed-run.mjs and tools/astra/speed-run-summary.mjs.

On every push and every pull request, the repository's continuous-integration workflow installs the locked dependencies on Node 24.22.0, runs the same site build Cloudflare runs, then runs the test files.

250 of the 375 test files run there.

The other 125 are listed with the reason they cannot: 111 need a real browser, 3 need Windows, 8 failed on the commit the list was measured on (26 October 2026), 3 for other recorded reasons. A test file in neither list fails the run, so a new test cannot be added and then silently skipped.

It references no secrets, has read-only access to the code, and deploys nothing.

The repository is private, so run results are not public. This page is generated from the workflow file and the test list themselves, so the figures above change when they do.

A change written with one company's AI model (Anthropic's Claude or OpenAI's GPT) is reviewed by the other company's model before release, and goes out only after that review signs it off with its findings fixed. These are automated reviews, not an external audit.

From 25 October 2026, 02:13 onward, 21 of the 23 releases in the record were signed off by a model from the other company before they went out. The two without one changed no file a visitor receives (test path fix; release branch carries the rate-limit release).

Before that, from 23 October 2026, 17:13, the log records a review for 16 of 39 releases, mostly those touching payments and orders, without saying which company's model wrote each change.

If any outside check or any byte check of the published files fails, the previous version is put back automatically and the outside checks are run again against it.

12 of the 62 releases in the record were rolled back automatically. After 10 of those rollbacks the outside checks passed straight away; after one, the first recovery run failed and a second passed; for one the log does not record a recovery run.

Text at least 4.5:1 in Light and Dark and 7:1 in High contrast; focus rings, input borders and button edges at least 3:1.

327 colour pairs (109 in Light, 109 in Dark, 109 in High contrast), computed from the site's colour tokens when this page was built.

Lowest text contrast: Light 4.94:1, Dark 6.54:1, High contrast 7.21:1. None below the floor.