I have been building a governance file for artificial intelligence for about a year and a half.
Twenty-seven protocols. A hundred and ninety-eight rules. It sits in a single document that an AI reads at the start of a session.
Yesterday I took it to a stranger.
Not a person. A different AI, on a free account, with no memory of me and no history with the work. It had never seen the file. It had no reason to be agreeable.
I loaded it and waited.
It refused.
Not the whole file. One part. The first part.
The file opened with something I had written months earlier. A gate. It said the session could not proceed until the AI had run the opening steps and produced its proof.
The stranger read that and said no.
It told me a document cannot switch itself on. It told me that a sentence sitting in a file has no power to stop anything. That the file was claiming an authority it did not have and could not have.
And it was right.
I want to be honest about what I felt. I did not like it. I had lived with that sentence for months. Four other cold sessions had read the same file and balked at the same spot, and each time I had explained it away. Wording problem. Reader problem. Not a real problem.
Four is a pattern. I had been calling it bad luck.
So here is what I did not do.
I did not talk the machine out of it. I could have. You can nearly always talk one of these into agreeing with you if you push hard enough. That is one of the things wrong with them.
And I did not cut the section out to make the objection go away. That is the other easy move. Delete the part that draws fire and the file reads clean.
Either one would have given me a document that passed. Neither would have given me a document that was true.
What I did instead was take the objection seriously and ask what the sentence could honestly say.
Here is what it says now.
The gate does not switch anything on. It sets an order. It says these steps come before the work, and if they did not run, the reader can see they did not run. Whether they run at all is the AI’s choice. The file’s only power is that the choice is visible.
That is a weaker claim. It is also the only one that survives a stranger reading it cold.
We changed a second thing while we were in there.
The file used to open with a compliance statement. The AI would say, in effect, I am following all of this. A pledge.
The stranger pointed out that a pledge is worthless. Nothing backs it. Any system can say the words and then do whatever it does.
So the pledge is gone. What replaced it is an exceptions list. The AI now states, at the open, every rule it is not holding for that session and why. If there are none, it says so.
Think about the difference.
A pledge cannot fail. It is just a sentence. An exceptions list can fail out loud, in front of you, in writing, at the start.
I gave up the reassuring version for the one that can be caught.
Now the part that matters, and it is not about my file.
This is what governance looks like when it is real.
Everybody in this industry publishes a framework. Principles. Commitments. Values statements with nice typography. What almost nobody publishes is the session where somebody with no stake in it read the thing and said this part is not true.
That session is the only evidence that the framework is anything but marketing.
A framework that has never been argued with has never been tested. It is a press release with numbered sections.
Mine got argued with by a machine that owed me nothing, and the argument made it smaller and more honest. Two of the claims I was proudest of turned out to be claims I could not support. They are gone. What is left, I can defend to anyone.
I will tell you the part that still sits with me.
The stranger was right and I was wrong, and I only found that out because I handed the work to something that had no reason to be kind to me.
Everyone I know is friendly to this project. Everyone in my own sessions has context, history, a reason to see it my way. That is comfortable and it is useless for finding out whether the thing works.
If you have built something, find the reader who owes you nothing.
Then do the hard part. Do not win the argument. Fix the file.
This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.
Post Library – Intelligent People Assume Nothing
Contact: micvicfaust@gmail.com
© 2026 The Faust Baseline LLC | All Rights Reserved
