I Write Rules, How AI Should Behave

There is a rule in the plumbing code I have been reading this week.

(Just follow me hear and see where this leads)

Your water service line and your house sewer have to run at least five feet apart. Horizontally. With undisturbed or compacted earth between them.

Not in the same trench. Not side by side to save a day of digging.

Five feet, and dirt in between.

There are two reasons for that rule, and a man who has laid pipe knows both of them.

The first is contamination. If one line fails you do not want it failing into the other one.

The second one is quieter, and it is the one I want to talk about.

You separate them so you can dig one up without disturbing the other.

If they share a trench, every repair is a repair to both. You cannot touch one without standing on the other. And thirty years later, when somebody opens that ground, he has no way to know which line he is looking at until he has already hit it.

Now let me tell you about a folder on my computer.

I write rules for how an AI is supposed to behave. I have been doing it for about a year. There are a good number of them now — seventeen protocols on one side, eight on the other, and rules under those running into the hundreds.

All of that could sit in one file.

It would be easier. One file to load. One file to hand somebody. One thing to keep track of.

It is six files instead. And every time somebody looks at that folder, the first question is why.

Here is the answer.

Different people are supposed to act.

One of my files is written to a machine. It says what the AI does — how it handles a claim, where it has to name its evidence, when it has to say it does not know.

Another one is not written to a machine at all. It is written to whoever runs the machine. It says what the operator owes the thing he is operating.

Those cannot live in the same file. An AI cannot obey a rule addressed to a builder. A builder cannot obey a rule addressed to an AI.

Put them together and you get a document where nobody knows whose job any line is. Which means when a line gets skipped, everybody can point at somebody else.

Separate files means separate answers to a simple question: who was supposed to do this?

And a draft can wear a signature.

This is the real one.

When I sign off on a rule, I read it in full and I put my name on it. That signature covers specific text on a specific day.

But rules do not all arrive at once. At any hour, some of what I have written is signed and some of it is still being argued over.

Now imagine both kinds sitting in one file.

That file has my name on it. It has a ratification block at the bottom. Anybody opening it sees a signed document.

And half of it is not signed.

The signature does not cover the new material. But it looks like it does. It is in the same file, in the same format, under the same name.

That is not lying, exactly. It is worse. It is a thing that lies on its own, without anybody deciding to.

So the unsigned material lives in its own file. It is called Pending Seats. It says right at the top that nothing in it is ratified. It cannot borrow authority from the signed work, because it is not standing in the same hole.

Then there is the fingerprint.

Every one of those files has a hash. A long string of characters computed from the exact bytes of the file. Change one comma and the string changes completely.

Six files means six fingerprints.

So when I open a session, I can compute all six and show that every one of them is the file it claims to be. If I changed one yesterday, five of them prove they did not move.

In one big file there is only one fingerprint. It changes every time anything changes. And it cannot tell you what changed, or where.

Now the part most people leave out.

The split costs something. I am going to say what it cost me.

Six files is six things to keep in agreement. And they do not stay in agreement by themselves.

Last week I found one rule written out in full in two different files. Same words, two homes. That means two places for it to drift, and eventually one of them says something the other one does not.

I caught it. I fixed it. One file now holds the words and the other one just points at it.

But that mistake only exists because I split the files. A single document could not have made it.

And yesterday I lost the folder. All of it. I got it back only because the files happened to be sitting open in a session. That was luck, not a plan.

So the honest version is this. Separation costs you bookkeeping, and it hands you a kind of error a monolith cannot make.

I still think it is right.

Because the error the single file makes is the one you never find. It does not show up as a contradiction. It shows up as a document that looks more settled than it is, sitting there being trusted, for years.

Five feet apart. Undisturbed earth in between.

You can open one without standing on the other. And a man who comes along thirty years from now can tell which line he is looking at before he puts a shovel through it.

” Attic Thoughts”-library – Intelligent People Assume Nothing

This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.