Zero ticks, on purpose

I run a public directory of wood-fired saunas in the UK — 279 operators, 68 builders, 15 guides, all researched rather than scraped. It is a real product with real readers, and it is also a deliberate experiment: how far can a publication be automated before judgement degrades, and what has to stay human for the answer to be "quite far"? The site says this about itself, in public, under my name.

The division of labour has held for months. Agents do the research-shaped work: discovering operators, filling missing fields, checking whether a website is still live, a booking link still books, a business still trades. A human — me — owns every write. Nothing an agent finds goes into the database on its own authority. Agent output is a queue of candidates with evidence attached, and each one gets approved, rejected, or corrected by hand. On a pilot run this month, the verification agent raised 77 flags across 20 listings; the flags took minutes of machine time and the review took longer, which is the right way round for a site whose claims carry my name.

The part I would defend hardest is the trust mechanics, because they are what stop "AI-run" from meaning "confidently stale". Three examples.

Every listing shows the date it was last checked, and the green tick appears only while that date is under four months old. When I set that threshold honestly this August, the entire corpus — checked in March and April — went tickless overnight. The site currently shows zero verified ticks. A softer threshold would have kept every tick lit. It would also have made the tick meaningless.

The monitoring assumes silence is the failure. A daily agent reads Search Console, the inbound enquiry queue, and — by sending itself a real email through the site's own route — whether mail actually works. Any failure writes a visible FAILED row into a log a health check reads at every session; a run that quietly does nothing is treated as the worst outcome, because it is the one that looks like health. This caught a genuine fault this week: a scheduled job dying at 09:15 each morning because it fired before the laptop's Wi-Fi was up, writing failure rows that read as site faults until the timestamps said otherwise.

And the cheapest check runs before any paid one. Before an agent spends tokens verifying a listing against the web, a zero-cost pass asks whether the record contradicts itself — whether the postcode field disagrees with the postcode sitting in the row's own address. That pass found 48 internal contradictions across 347 records this week, including two postcodes that four months of externally-focused checking had filed as unfixable. Most data errors don't need research to find. They need someone to ask the row what it already knows.

What stays human is narrow and named: database writes, editorial voice, anything that becomes a public claim, and the decisions about what the site will not do — no invented listing data, no pay-for-placement, no health overclaiming. The site's AI policy page states which tools do what, and the editorial machinery behind the guides — a multi-model research pipeline where every layer catches the previous layer's errors — has its own write-up.

The ambition, stated plainly on the site itself, is a largely automated publication with a human in the loop checking everything. Automation does the checking at a scale I never could: re-verifying the full corpus costs roughly £55 in machine time, a region-sized batch each week for about fifteen weeks. The human gate means a reader never encounters a fact that only a model believed. And the ticks are the visible edge of it. They come back region by region between now and the winter, each one carrying the date it was earned — which is the only form of "verified" I would accept from anyone else's directory, so it is the only form this one shows.