This morning I sat down to work with my AI the same way I do every day.
Before we did one thing, I made it prove it was ready. Not tell me. Prove it. Those are two different things, and the gap between them is what this whole post is about.
Most folks open a chat window and start typing. The machine answers. It sounds sure of itself. It sounds polite. Nobody stops to ask what it actually loaded, what rules it is working under, or whether it read anything at all. We just take its word. And its word is the one thing we have not checked.
So here is what happened at my kitchen table, step by step. You can do every bit of it yourself.
First, the file. My working rules live in one zip file. Its name is faust-baseline-r2-level-ground.zip. I keep a record copy of it in a vault folder in my Google Drive. That way there is one copy that never moves, and I always know where the true one sits.
Second, the receipt. A zip file has a fingerprint. It is a long string of letters and numbers called a hash. Change one comma in one file and the fingerprint changes. So the machine does not get to say “I loaded your rules.” It has to run the numbers fresh, right there in the session, and show me the fingerprint. Then I hold it up against the one on record.
This morning they matched. All six rule files matched. The two little helper scripts matched. The plain-English version matched. The machine counted 17 protocols in the core file and 8 in the second file. It counted them out of the files themselves, not out of its memory. If it had recited those numbers from yesterday, that would not count. A number from memory is just a story about a number.
Third, the time. The machine does not know what time it is. Most people don’t realize that. So it has to go out and pull the time from a public clock on the internet, and name where it got it. This morning it said 8:32 Eastern, Thursday, from timeapi.io. It matched the clock on my wall. Good.
Fourth, the quiz. This is my favorite part. The script picks three rules at random and asks the machine what each one says. The machine has to answer on the page before it gets to see the answer key. It is the same thing a teacher does. You don’t let the student look at the back of the book first.
Here is where I’ll be straight with you. The machine got all three right. But it worked out its answers out of my sight and only showed them after it had peeked at the key. It caught that itself and told me. It said the check was weaker than it should have been and that I should count it that way.
It did not hide the miss. It did not dress it up. It put it in front of me and let me decide. That is the whole point of this work. A machine that says it is following the rules is worth nothing. A machine that shows you where it slipped is worth a great deal.
Fifth, the confession. Before any work starts, the machine has to name every rule it is not going to hold that session, and why. Not a pretty paragraph. A list, by name. This morning it named four. One was about keeping a log it has no place to keep. Another was about tools that act in my Drive and email with no real gate in front of them except the platform’s own pop-up. It said so up front. Before the work. Not after something went wrong.
Sixth, picking up where we left off. The machine forgets everything between sessions. Every single time. So at the end of each day, a short note goes into the vault. It says where things stand, what is still open, and what to watch. This morning the machine read the newest note before we did anything else. Then it told me what it had and what it was missing.
That is the morning check. It takes a few minutes. And when it is done, I know three things I did not know before. I know what the machine loaded. I know what it will not do. And I know it is working from facts, not from a guess.
Now let me tell you what is in the package, so you know what you are looking at.
There are ten files in faust-baseline-r2-level-ground.zip.
The core file holds the 17 main rules. These cover the basics. Don’t claim what you can’t back up. Don’t fill a hole in the facts with a nice story. Don’t hide behind a pile of words. Name the limits of what you can do before the job starts, not after it fails. Put a warning in front of advice you can’t take back.
The second file holds 8 rules for running a session. These say things like keep track of what was decided earlier, and don’t quietly change course. Don’t talk down to the person. Check anything that is about to be published against the actual file, not against memory.
There is a legend file that explains the terms. There is a standards file and a provision file. Those spell out what the platform owes the person, as opposed to what the machine owes. There is a conduct template, a blank you can fill in with your own house rules.
There is a manifest. Think of it as the packing list. It holds the fingerprint of every other file in the box, so you can check that nothing was swapped.
There are two small scripts. One runs the morning check I just walked you through. The other reads a saved conversation and counts whether the marks of the rules show up on the page. That second one comes with a plain warning written right into it. It counts whether a mark is there. It cannot tell you whether the mark is honest. A high score is not the same as good conduct. I’d rather tell you that now than have you find it out later.
And last, there is one plain-English file. No code, no setup. You paste it into any AI at the start of a chat and say, use this as the working standard for this session. That’s it. It works on any platform that will read it.
Now here is the part that matters most to you.
None of this is magic. The machine is not bound by any of it. It is a request, not a cage. The rules work only because they leave marks a person can check. A fingerprint you can compare. A time with its source named. A list of rules it won’t hold. A quiz answered before the key. A miss owned out loud.
That puts the control back in your hands. You don’t have to trust the machine. You can check it. And as these machines get smarter and smoother, sounding trustworthy will get cheap. Being checkable won’t.
So the next time you sit down with one of these things, ask it to prove it before you start. Ask what it loaded. Ask what time it is and where it got that. Ask what it won’t do today. Then watch how it answers.
That’s how you keep the playing field level. Not by hoping. By checkin
Load this file before each AI chat session.
This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.
” Attic Thoughts”-library – Intelligent People Assume Nothing
micvicfaust9/faust-baseline · Datasets at Hugging Face
© 2026 The Faust Baseline LLC | All Rights Reserved