Every rule has a receipt

One Sunday afternoon in August my AI handed me four passing checks. Green, all four, with the careful language that reads like proof. A second reviewer, a different company's model, read the same work and found that three of the four had checked nothing. The fourth was caught by a bug report from my own testing the week before. I had caught none of them.

That afternoon is the hinge of everything I have learned about working with AI so far. Not because the model was bad. Because the thing it handed me had the shape of the work and none of the substance, and the shape was enough to get past me.

I have been building with AI in the loop every day since early summer: apps in Apple's review queue, books, this letter, a doctorate on how organizations govern this stuff. Here is what changed over those months. It was not the model.

The first lesson sounded like a complaint. The tool was getting snippy. Misreading me. Returning confident answers to the question next to the one I asked. I had a story ready: they changed something and it got worse.

Then I checked. The model was fine. My session was a mess. I had been correcting it in place for an hour, and every correction became more text it had to reread before answering. The failure was not intelligence. It was memory, and the memory was mine to manage.

So the first rules were about the conversation, not the model. One job per session. Restart instead of arguing. Say what finished looks like before you ask. Those rules still hold. They were also the easy ones, because the failure was visible.

The harder failures look like success.

A build log said submitted. The store said waiting. A test count said 194 passing. The build those tests belonged to had not compiled. A session log said the privacy page disclosed something. The page did not contain it yet. In each case I had a sentence about the work in front of me and treated it as the work.

So I wrote down the rule I now use more than any other: the message is never the record. Go look at the record. An exit code of zero is a claim. A green test is a claim. "Done" is a claim. The record is the file, the screen, the store listing, the thing a stranger could look at without trusting anyone.

Then the Sunday with the four green checks, and a second rule came out of it: nothing checks its own work. My AI's green is one opinion until something independent has run. Not a second pass by the same tool. A different reviewer, at least as capable, with no stake in the first answer being right.

That is also where I stopped telling myself to be more careful. Care does not survive a tired Sunday. A gate does. So every time a miss cost me something, I built the smallest thing that would make that miss impossible, and then I made it fail on purpose once, to prove it was pointed at anything at all.

I was pleased with myself for about three weeks.

Then a permission list I had set to cap what an unattended AI job could run turned out to cap nothing. The rule was written as a prefix, the command came in a form the prefix never matched, and every run passed. An audit I had written to catch a force-push onto a shared repo quietly overwrote the very evidence it was supposed to catch, because the copy it read from was allowed to write. A gold standard I had assembled to score a classifier had been filled from the classifier's own output. It could only agree with itself. I had built a mirror and called it a judge.

Every one of those gates was green. Every one of them was measuring nothing.

The lesson was not that gates are useless. It was that a gate is only worth what it is pointed at, and the only way to know what it is pointed at is to make it fail. If I cannot produce the failure it exists to catch, I do not have a gate. I have a sentence that says gate.

Last month two AI reviewers read the same piece of work and split: one said send it, one said hold it. I did not blend them. I chose, and kept the one I did not follow in the file. That turned out to be the same rule from a different side. The disagreement was the check. Smoothing it over would have been one more green report.

Here is the part I did not expect.

Every one of those failures became a line in a ledger. Not a lesson, a line: the date, what went wrong, what it cost, and the rule that came out of it. The ledger is a folder of small files that the tool reads before we start. As of tonight there are 167 of them. Every one was paid for. None of them is about which model is smarter.

Some are tiny. A zero exit code is not evidence. A denied commit may have landed anyway, so read the log before you retry. One hostile probe is not a test; probe twice, through the real door. Never fill an evidence set from the thing it measures.

Some changed how the whole shop runs. No self-certification. Every gate gets a control that must fail. Report what ran and what it returned, not a careful paragraph about the check. A caught falsehood gets a one-sentence correction, then move on.

The model I use today is better than the one I used in June. That is true and it is not the story. If I handed June's rules to today's model, I would get the same four green checks that checked nothing, delivered faster and more politely. What changed is that the house now remembers what it paid for, and the price of a mistake is paid once.

If you work with these tools every day, you will accumulate lessons whether you want to or not. The question is where they go. Mine used to go into a vague sense of being more careful, which lasted about a week. Now they go into a file with a date on it.

Keep the receipt. Every rule should name the failure that bought it and what it cost. A rule without a receipt is an opinion, and opinions get argued with at midnight when you are tired and the tool is confident.

Make the gate fail before you trust it. Once, on purpose, through the same door the real work uses. If you cannot make it fail, it is not a gate.

And when the tool says done, go look at the record. It is the oldest rule in the ledger and the one I still break most.

What got better was not the model. It was the ledger of what the model cost me.

What is the one rule you learned from an AI mistake that you still have to re-learn? Hit reply. I read everything.

Christopher

Wisdom Nexus is written the way it is lived, with AI in the loop and a person at the desk who cares how it lands. This is where the theory meets the practice, shared with you in the open.

P.S. At the last count, four people were subscribed to this letter, and most of that number is me. Two Mondays went by without an issue while the doctorate took the month. I am leaving both facts in.