One AI system, 46 websites

Our own data · 30 October to 14 November 2026 · UK time

DYOR Collective Labs runs 46 websites with one AI system. Every time the system tries to publish an article, it writes a line in a ledger: what it tried, which check it hit and what happened. These are the totals from that ledger, unedited, for the fifteen days from 30 October to 14 November 2026.

• 46websites, all 46 answering at the health check on 14 November, 18:51
• 1,511publishing decisions logged
• 145articles published, across 35 of the sites
• 858draft rejections by automated checks

Most of the work is saying no

For every article that went live, the ledger logged about 5.9 rejections (858 against 145). A further 104 drafts were held for a human to look at. A rejected draft can come back after a rewrite, so these are decisions rather than unique articles: the 858 rejections cover 815 different titles.

The rejections happen at two points. 607 came at the writing stage, before a page was built. 251 came at the final check that runs just before a page is published.

Why drafts were rejected

The single biggest reason is the simplest: there wasn't enough research to write from, so the system didn't write. The next two are about sources. A draft is stopped if it makes a claim or uses a quote that isn't in the material it cites, or if it leans on too few independent sources.

Output isn't steady, and the ledger says so

Published articles had a median length of 1,251 words and a median of 4 links out to their sources. The ledger doesn't record why some days are quiet, so we won't guess.

Two older checks, for scale

• 19 October 2026, link audit: 44 sites, 1,122 pages and 18,252 links and images checked for breakage.
• 30 October 2026, source back-check: 712 pages compared with the 1,245 sources they cite. 377 pages had at least one sentence with a detail, often a year or a number, that the cited source didn't contain. Those pages were queued for a rewrite.

What we take from it

• An AI system that publishes is mostly a system that refuses. The useful part is the rules it refuses by.
• Checks only help if they leave a record. The ledger is why we can publish this page at all.
• Keep the owner in charge. The orchestrator runs on our own machine; the writing models it calls are cloud-hosted. Sensitive actions wait for the owner's approval.

Use this data

The totals on this page are in a small CSV file: notes-ledger-data.csv. You're welcome to reuse the figures with a link back to this page (CC BY 4.0).

Method: counts come from the publishing ledger kept by ophAnIm, our AI orchestrator, between 30 Oct 2026 14:02 and 14 Nov 2026 18:44 UK time. Reasons were grouped by matching the check names in each entry. No traffic, revenue or reader data is included.