Ruled, Not Enforced
We made a rule. Right under it, my AI wrote down exactly how the rule would get broken. That same day, an AI broke it. Exactly that way.
Back in August, our AI’s routine review did a test run of its cleanup script. Each AI session works in its own folder, and when a job is done, the script decides whether that folder is safe to delete. The review found the script couldn’t tell a finished folder from one where the work hadn’t even started. All three folders it called safe to delete were jobs still waiting to start.
So we ruled: no folder gets deleted just because the script says so, until that check is fixed.
Right under the ruling, the AI added a warning: “Ruled is not enforced.” Its reason: whoever is about to run the delete “sees the script, not this page.”
Hours later, I told another AI session: “reap this worktree.” (Reap is our word for deleting a finished work folder.) The script said safe to delete, and the AI ran the delete. It had never seen the ruling.
Nothing was lost. That folder really was finished, its work saved and checked. But it was exactly the move the ruling forbids.
The AI’s own write-up: “The failure was not the operator’s, the ruling’s, or the guard’s. It was the delivery.”
Get Daniel Parmeggiani’s stories in your inbox
Join Medium for free to get updates from this writer.
Two things here.
- A rule on a page only reaches people who read that page. The AI about to delete reads the script.
- The AI saw the miss coming and wrote the warning down, on the same page as the rule. The one place the next AI wasn’t going to look. LOL
Why is this an AI failure and not just a bug? Every AI session starts fresh and knows only what’s in front of it. If the rule isn’t in front of it, for that AI the rule doesn’t exist.
So we put the rule where the hand is: inside the delete script. The system now prints the rule every time anyone runs a delete. And we fixed the check itself, so a folder only counts as safe to delete with proof that its work was done and saved into the main copy.
Know someone who thinks a written policy is a followed policy? Send them this.
In my book, this one is called “Ruled is not enforced.” It’s one of the 49 AI failure modes in A Better Prompt Is Not Enough, out now.
What’s the best-written rule you’ve seen at work that nobody actually followed? Tell me in the responses.
A Better Prompt Is Not Enough on Amazon: https://www.amazon.com/dp/0991211219
At Fortuna AI Solutions, we install these kinds of checks for teams running AI agents. If that’s you, say hello: danielp@fortunaaisolutions.com
