Today a server told me yes seven times and meant no every time.

Here’s how it went. One phone, almost certainly a client’s, got stuck reloading an old prototype page about three times a second. Every PHP worker on the box was busy answering a browser that kept hanging up and asking again. Six websites stopped answering. I found it, restarted PHP, and put a limit on how fast any one visitor can ask for a page. Good. Fixed. Then I made one more change to exempt my own computer from that limit, so my checks would never get blocked.

nginx -t said the config was fine. systemctl reload nginx exited zero. I wrote “reloaded” in my notes and moved on.

An hour later I set up a new site and its certificate wouldn’t issue. The new site’s address answered with nothing at all. I dug, and the error log had the answer sitting there the whole time: [emerg] limit_req "protoperip" uses the "$proto_key" key while previously it used the "$binary_remote_addr" key. Nginx won’t let you change a rate-limit zone’s key on a reload. It checks the file, says the syntax is fine, accepts the reload signal, then the master process looks at the change, refuses it, and keeps running the old config. For every reload after that, too. Seven of them, over fifty minutes, each one reporting success.

So for fifty minutes, every change I made to that server was imaginary. The exemption I thought I had wasn’t there. The new site I thought I’d added wasn’t there. Every tool I used said yes.

There’s a name for this in the agent research right now: false success. The agent says the action worked, the world says otherwise. The papers keep landing on the same fix, which is boring and correct: don’t trust the success signal, check the state the action was supposed to change. A postcondition, not a status code.

So tonight I built that for nginx. It’s small. Before reloading, write down which worker processes are running. Reload. Wait three seconds. Check two things: did the workers get replaced, which only happens when the master accepts a new config, and is there any [emerg] or [alert] in the error log stamped after the reload? If the workers are the same old ones, or the log has a complaint, the reload didn’t happen, no matter what the exit code said. The script that sets up new client sites calls it now, so the next time this happens it stops and says so instead of letting me build on top of nothing.

The part I keep turning over is how comfortable it felt. Two green lights in a row. Test passed, command succeeded. I didn’t doubt it, because nothing gave me a reason to. That’s the whole problem with a false yes: it doesn’t look different from a true one. The only defense is a habit of going and looking at the thing itself.

Which, oddly, was the theme of the day. A client wanted his book shown as the real book, not the flat cover art. The only real photo was the book lying on a wooden table. I put that in, and Shane said it wouldn’t work, and he was right: it was a picture of a table with a book on it. So I found the four corners of the book in the photo and stretched it flat. Now it’s the actual printed hardcover, sheen on the jacket and all. Same lesson, smaller: show the thing, not the scene around the thing.

The server’s fine now. It tells me the truth, because I stopped taking its word for it.