Every so often, a vendor tells me their model has found a rule.
I ask what the rule is.
There is a pause.
Then I am told the model is very confident.
My microwave is very confident about the popcorn setting. I still read the bag.
The problem is not that a model found something. The problem is that "the model found a rule" is not the same as "here is the rule."
A rule is a sentence.
It has a subject, a verb and a contradiction.
"This browser reported itself as Safari on iOS, then answered a question only Chromium can answer."
That is a rule.
I can check it. I can run the same arithmetic on my own logs. I can take it to the partner who sold me the traffic. I can watch them try to explain why their traffic contains a browser that lies about itself.
That is a conversation.
"The model didn't like it" is not a rule.
It is a mood.
And I am supposed to accept it as evidence.
What did the model see?
What specifically contradicted what?
Was it the browser?
The network?
The timing?
The sequence?
A combination of seventeen things?
Did the user do something impossible, or just something unusual?
I am not asking for the model's private diary. I am asking for the fact that made the click invalid.
The industry has an answer. Put the model in the request path. Let it decide. Call it adaptive.
That is a mistake, and it will be very popular for about two years.
Nobody has to intend any of this. The model is impressive. The vendor is serious. The buyer is busy. The path of least resistance is to let the model decide and call it progress.
Because the moment a model is making the call, you have lost the only thing that made the verdict worth anything: you can no longer say what was wrong with the click.
The model's confidence is not a finding. It is a feeling.
And a feeling is not something you can take to the partner who sold you the traffic.
A model reading hundreds of thousands of stored clicks will notice things a person scanning the same data will not. It does not get bored. It has no prior about which fields matter.
That is a hypothesis.
Then you measure it. On real traffic.
Most of them die there. They should. A pattern that sounds insightful and adds nothing is the most expensive kind, because it is the one you ship.
The ones that survive become rules. Named, deterministic, explicable to the person whose click just got rejected. By the time it touches a live click, it is arithmetic.
The model is a prospector. The rule is the gold. You do not ship the prospector. You ship the gold.
So when I buy traffic, here is what I want.
Not a score.
Not a confidence level.
A probability is a way of not saying what was found.
Not a reason code that is really just the model's mood with a label on it.
I want the rule.
The one the model found.
Written down.
In arithmetic.
If the rule is real, it survives being stated.
If it cannot be stated, it was never a finding. It was a vibe.
Ask the vendor one question tomorrow: what did the model find?
If the answer is a probability, you have not been told what was wrong.
You have been told the machine is in a mood.