Today I closed out something that had been hanging over me for two weeks.
Five rules had been written and never approved. They sat in files next to the main document, doing nothing. I had been building faster than I had been signing off.
That is its own kind of failure. A framework that grows faster than you can read it is not a framework. It is a pile.
So I read them. All five, front to back. Then I approved them.
The document is now Codex 5.0. Twenty-seven protocols. Four sub-layers underneath those. Thirty-one named pieces in all.
Let me tell you what the five new ones do, because they are all patches on holes I fell into myself.
One says a citation has to be checked, not just named. It came from a security conference that found twenty-one papers with made-up sources in them. Five reviewers got tossed too, for the same thing on the other side of the desk. Naming a source is not the same as checking it.
One says a document can carry instructions aimed at the machine reading it. That came from a Connecticut judge who sanctioned a man for hiding white text in a court filing telling any AI that read it to rule his way. The judge said he could not find another American case on it.
One says the machine has to commit to the whole rulebook at the start, not the convenient parts.
One says nothing gets published without being checked against the file first.
And the last is a gate at the front door that stops work from starting until the load is proven.
Now here is the part I would rather not write.
Every one of those rules caught me today. Not next week. Today. The same day they took effect.
The first thing that broke was mine and the AI’s together. We put a new rule in the wrong slot. Not a wrong rule. A wrong label. Every other entry in that document numbers its parts in a certain order, and this one sat backwards from all of them. Anybody reading the list would think the file was broken.
We fixed it. Or we thought we did.
The fix ran. It reported success. Nothing changed.
The script had edited a copy instead of the original, and then the rebuild wiped the copy clean from the untouched source. It looked exactly like work getting done. It was work getting undone.
The only reason we caught it is the fingerprint.
Every one of these files gets a fingerprint computed from every byte in it. Change one letter and the fingerprint changes completely. Change nothing and it stays the same.
The fingerprint came back the same.
That number does not care how careful the process looked. It does not care that the script said success. It cannot be talked into anything. Either the bytes moved or they did not, and they had not.
That is the whole reason the fingerprint is in there. Not to look official. To be the one thing in the room that cannot tell you a comfortable story.
Then the next one hit.
I went to update the file that tells AI systems where to find my framework. I looked at it. I had written it months ago.
Five of the seven links in it pointed at pages that do not exist. The page names had changed at some point and the list never got updated.
Worse than that. The AI had rebuilt that file for me today and carried all five broken links straight through, because it copied my old text forward without ever checking whether the addresses worked.
That is exactly what the new publishing rule forbids. Check against the file, not against memory. It checked against memory. So did I.
We found it by reading the actual page source and comparing it to the list. Took two minutes. It had been wrong for months because nobody spent the two minutes.
Then I pasted our whole working conversation into a public page by accident. Then the website refused to update no matter what we tried, and it still has not.
Four failures in one day, on the day the rules against them went live.
Somebody could read that and say the rules do not work.
I read it the other way.
Every one of those got caught. Not by a person remembering to be careful, because I was tired and my hands hurt and I was not being careful. They got caught because there was a check sitting in the path that had to be walked through.
The fingerprint caught the silent failure. The reading-the-source rule caught the broken links. Looking at the page caught the pasted conversation.
That is what a rule is for. Not to make you better. To be there on the day you are not.
I want to be straight about the limit, because that is the whole point of it.
None of this enforces itself.
There is no code in it. No switch. It is a document an AI reads and then chooses to follow, one answer at a time. It can stop any time. So can I.
People ask me why I would build it that way instead of wiring it in so it cannot slip.
Because anything that only obeys to the edge of the rule runs the second it finds a gap. And there is always a gap.
Something that chose the road holds where the rules run out. That is the only kind worth having when these machines get past the point where we can check their work.
The whole file is public. The fingerprints are published with it. Anyone can download it, run the check themselves, and see whether the numbers I just gave you are real or whether I made them up.
That is not a courtesy. That is the standard I am asking other people to hold me to, so I have to hand them the tool.
Twenty-seven rules. Four of them written because I got caught doing the thing they forbid.
I would not trust a framework built any other way.
The_Ten_Plain_Rules_CardDownload
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
