You buy traffic. Some of it gets blocked. You ask why. The vendor tells you the click was invalid. You ask what that means. They say the model flagged it. You ask what the model saw. They give you a number: 73% confidence.
Excellent.
You now have a probability. You do not have a reason.
That is the moment the conversation ends, and it is the moment you should start asking harder questions.
A model in the request path makes a call. That call comes wrapped in a confidence score. The score feels scientific. It has decimal places. It is a horoscope with a decimal point. It suggests the machine has weighed the evidence and reached a considered view.
But a probability is not evidence. A probability is a way of not saying what you found.
You cannot audit a probability. You cannot challenge it. You cannot take it to the partner who sold you the traffic and say, "Explain this." Because the partner will ask what the model saw, and you will not know. The vendor will say the threshold was your setting. And every wrong call becomes an argument about your configuration.
That is a structural consequence, not a plot. Nobody has to intend it. The moment the reasoning stays on the vendor's side, the customer cannot check it. Ambiguity is not neutral. It flows toward the party that controls the definition.
Now consider what a model is actually good at.
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. It can find a pattern: these supposedly unrelated users share a timing signature, a device combination, a contradiction in the browser story.
That is a hypothesis.
Then you measure it. On real traffic. Most hypotheses 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. By the time a rule touches a live click, it is arithmetic.
"The model didn't like it" is not something you can take to the partner who sold you the traffic.
"This browser reported itself as Safari on iOS, then answered a question only Chromium can answer" is something you can take to the partner. You can watch them try to explain it. That is the difference between a verdict and a fact.
The industry's answer to the acting bot is to put a model in the request path and call it adaptive. That is a mistake, and it will be very popular for about two years.
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 didn't like it. That is not a reason. That is a mood.
Adding a confidence score does not fix this. It makes it worse. "Invalid, 73% confidence" is not more transparent than "invalid." It is less. You have now been handed a number you cannot audit, cannot challenge, and cannot take anywhere. And the vendor has acquired a permanent defence: you set the threshold.
So the demand should not be "give me a confidence level."
The demand should be: What did you observe, and why does it mean what you say it means?
What signals were observed? What specifically contradicted what? Can I check the reasoning myself? Can I take it to the counterparty and have the argument?
If the answer is a probability, ask again. The model found a pattern. That is useful. The model made a verdict. That is not.
We think the right division of labour is not subtle. Use the models to find the pattern. Do not use them to make the call.
That is the whole of it. A buyer should not have to argue with a number. They should be able to argue with a fact.
Ask your vendor for the arithmetic. If they hand you a probability, ask again.