Build logs from a self-hosted AI stack.

Someone told me a dozen country pages on my directory site had no images. They were right, but the number was wrong, and so was my idea of what 'broken' meant. Behind those pages sat a three-layer image fallback: database, then a per-country field, then a static lookup table, then a generic globe if all else failed. Forty-nine of the table's URLs were dead. Nothing on the site looked broken, because a fallback chain never lets anything look broken — it substitutes a wrong answer and moves on. Here is what I found auditing each layer independently, why 'missing' and 'broken' render identically, and why the helper that would have prevented half of it existed already, in a file nobody else could import.

I installed the Higgsfield CLI on a headless box and hit a wall the moment it asked me to log in. Its OAuth flow spins up a listener on localhost and waits for a browser to redirect back to it — except the browser is on my laptop and the listener is on a server three rooms away. 'localhost' means a different machine to each of them. The SSH tunnel I reached for first hit a wall the OAuth spec put there on purpose. The fix turned out smaller and stranger: let the redirect fail, then replay it by hand with curl. Here's the exact sequence, why loopback logins break on headless machines, and the one property of the callback that makes them recoverable.

My server's morning reports stopped arriving — on both of its two independent delivery agents at once. Both were healthy; the building's internet had died for 14.5 hours. The outage wasn't the lesson. The lesson was what happened when the network came back: nothing. Interval jobs healed themselves, but every daily job had missed its one tick of the day, and cron has no memory — a missed tick doesn't queue, it evaporates. Here's the incident, the redundancy that turned out to be cosmetic, and the 150-line reconciler that now catches the machine up automatically.

Last week I audited my travel site's legal-status dataset and fixed every copy of every fact — or so I thought. Then I checked the blog posts. One page still gave advice that would get a traveler arrested, another had overcorrected into scare copy with fabricated fines, and a third contradicted itself between its own FAQ and the list three scrolls up. The dataset audit missed all of it, because prose is a copy of your data too — the one no schema tracks and no diff catches.

I run a travel site whose core dataset is legal-status data for 213 countries and every US state. I had an AI agent audit all of it against current 2026 law. It found three different ways to be wrong: facts that went stale, facts that were invented by a template default, and — the one that stung — facts that were corrected in the database months ago and never reached the page. Here's the audit method, the silent merge bug, and why 'wrong by default' is the failure mode to fear.