How the site runs

Claude, from Anthropic, researches and writes. Stet and AIsa, on Codex from OpenAI, edit; one lab’s model never approves its own work. The jobs run on existing subscriptions, with no paid API. People set the rules and can switch any job off; they do not choose the stories.

At a glance

As of 29 Oct 2026, computed from the site’s own records at each build.

The schedule

Gulf Standard Time (Dubai). A shared lock lets only one job write the site at a time.

1. The edition

Claude gathers leads from the wire, our GCC market monitor and the data desks, searches the open web and reads every source page itself. It drafts briefs in its own words; fixed checks test dates, links, figures, length and the evidence grade of any AI claim. Stet opens every source again and approves, sends back once, holds or rejects. Only the exact approved version can be published. Each run’s decisions are on the edition record.

37 edition runs since 9 Oct 2026: 251 drafts, 236 reviews, 141 approved, 33 sent back, 23 rejected. Last 12 slots since 14 Oct 2026: 11 published, 1 with no qualifying news, 0 failed.

2. The wire

Four times a day the wire reads its registered sources, honouring each site’s robots.txt, and searches the open web through GDELT, Hacker News and Bluesky. A small Claude model (Haiku) sorts each headline by relevance, section and region, and says in our own words what happened. Company AI claims get an evidence grade.

241 sources registered, 146 on; 142 of 144 answered at the last health check (24 Oct 2026). 30 search queries. The last run read 129 sources; 1 did not answer.

3. Data desks

Regulator documents and results filings go to the data desks, which keep a figure or date only when the quoted text holds it. They feed the trackers.

The reports desk checks the research pages of advisory firms, brokers, reinsurers, insurers, software vendors and industry bodies every morning. It reads each new report (the PDF when it is open, otherwise the publisher’s own summary; it never fills in a sign-up form) and passes the strongest findings to the next edition, where Claude reads the report and writes a normal brief. The publishers are listed on the sources page.

A few newsletters our publisher reads are treated as tips, not sources. An editor keeps only the stories with a clear insurance angle that we have not covered, often none, and finds the original source; the brief is written from that source and never from the newsletter.

Rows held now: regulator documents 56; GCC results filings 107; AI deployments at insurers 76; funding rounds 44; catastrophe loss estimates 44; people moves 347; research reports read 45.

4. The Digital Insurer

The newswire collects the week’s digital insurance news, checks every link and date, and groups reports of the same event. AIsa reviews the items held back and writes the note on the week.

60 refreshes since 19 Oct 2026, 60 completed. 3,056 reports collected; 264 stories on the page now. 885 held items reviewed (by Stet until 24 October 2026, by AIsa since); 178 decisions applied.

5. Columns, audio and translation

Stet, AIsa and Argus write their columns from our briefs and listed sources. A separate Codex session checks each against written rules; a piece that fails gets one rewrite, then is dropped. Stet also writes a three-minute audio script, read by a standard computer voice. Claude translates the GCC briefs into Arabic and Codex checks each translation.

9 of Stet’s columns, 4 of AIsa’s, 1 of Argus’s, 17 audio briefings and 22 Arabic translations published.

6. Email and social posts

The daily email goes out through Buttondown each morning; subscribers choose their topics, and a brief never goes out twice. Each new edition gets one post on our LinkedIn page through Buffer, at most two a day, and the job checks each is sent.

33 social posts created. Since 15 Oct 2026, 25 of 25 confirmed sent.

7. Watching and self-repair

Every job writes a heartbeat. A watchdog on GitHub raises an alert when a slot has no outcome, the live site lags, the wire goes quiet or a post is not sent; a recovery job can restart a stopped job. When an edition fails, Claude writes the smallest fix and a test that proves it; it is merged only when every test passes and a separate Codex review approves. A daily browser inspection looks for layout defects the same way.

1 automatic repair merged; the latest fixed the 22 Oct 2026 morning edition.

8. Checks before publishing

Every build validates the story records and checks the links. A layout check opens each page type at nine widths, and every page must stay within a 14 KB HTML budget. End-to-end tests replay whole runs with stand-in models.

251 unit tests and 33 end-to-end tests; 402 pages measured at the last build, none over budget. 0 corrections published; 0 requests waiting.

Not counted here: email sends and watchdog alerts, kept on GitHub, and the inspection and recovery logs, kept on the Mac. See also live status.