HOW IT RUNS
How this site runs
I run this site alone — or more precisely, with an AI team I built. Here's what runs on its own, what checks itself, and where I sign off. It's the same thing I did when running labs: decide first who does what, how it gets checked, and who makes the call.
What runs on its own
- Briefings. AI gathers the day’s signals and drafts a briefing, grading each source by strength of evidence. Only issues I approve go out to subscribers.
- Publishing, every morning. Pull → build → check → swap → stability test. If any check fails, the site stays exactly as it was the day before.
- English editions. The same system prepares them. Anything published under my name — like the handbook — is written separately for English readers, not translated.
What checks itself
- A pre-publish checker. Every publish, it looks for eight kinds of problems, including email addresses, location data left in photos, internal paths, development URLs, broken links within the site and drafts I decided not to publish. Before it checks anything, it has to catch eight problems I planted on purpose. A checker that can’t catch anything shouldn’t get to say “all clear.”
- A daily outside check. It opens the live site from outside: are public pages up, are private ones locked, are the certificate and security settings intact?
- A monthly cross-audit. A model from a different AI provider does a read-only review of the site and its operating records. I check whether each finding holds up before fixing it. A system built by one person tends to have blind spots that all point the same way, so I added a second pair of eyes on purpose.
- An error log. When an AI claim turns out to be wrong, it goes in the log — whether I caught it, another model found it, or later measurement exposed it. Across all my AI work, this site included, the log has passed entry 300. Those aren’t outages — they include small misreadings and unchecked claims. Counting mistakes is how a repeat mistake gets blocked by a mechanism instead of another rule.
Where I make the call
- Sending to subscribers, publishing and judging forecasts, and publishing handbook entries all need my approval.
- AI doesn’t decide what goes public.
- AI doesn’t write the rules either. It finds gaps and reports them; I decide how to close them.
Forecasts get logged, then judged
Every forecast is recorded with the condition and the date that will settle it. When that date comes, it’s judged a hit, a miss, or impossible to call on the forecast log. Once a forecast is published, I don’t edit the wording.
Where this setup falls short
- A checker only catches problems it knows about. That’s why the outside check and the cross-audit overlap with it.
- AI drafts can be wrong. That’s what the error log is for.
- This is a one-person system. An organization with more people would need to spread approval authority further.
To see what the team is working on right now, visit the AI team office.