The model is the witness, not the judge.
A lot of people can get an AI model to do something impressive once. Far fewer can get it to do the same thing ten thousand times without quietly lying to someone. The distance between those two is the whole job, and it comes down to a single rule I follow in everything I build. The model is a witness. It is never the judge.
A witness tells you what it saw. You do not hand a witness the gavel. You take what they say, you check it against the evidence, and you decide. That is exactly how I treat a language model. It reports what it observed. The verdict is computed afterward, in code, from the evidence it gave, and the model is never allowed to declare something true just because it sounded sure.
Here is what that looks like in practice. One of my products is a voice matchmaker. A person has a real spoken conversation with an AI, and out of that the system builds a profile of who they are. The model proposes what it heard. Then code checks every line before any of it is kept. The quote has to be the person's own words from their own turn. The reading has to be supported by that quote, or by the question it answered. If the model says someone has a dealbreaker, the code has to find where they actually confirmed it, or the claim is quietly downgraded. A sentence the person said as a negative, read back as a positive, is stored as uncertain, not as fact.
The model gets to speak. It does not get to be believed on its own say-so.
The same rule runs through INTELLaMiNER, my B2B platform, from the other direction. It surfaces a reason to call a buyer only when that reason traces to a real source. A signal it cannot stand behind is dropped before a rep ever sees it. Not softened, not flagged. Dropped.
Why this matters is simple. The most dangerous thing about a good model is how convincing it is when it is wrong. It does not hedge. It hands you a clean, confident, well-written falsehood in the same voice it uses for the truth. If your product treats the model's output as the answer, you have built something that is right most of the time and has no idea when it is not. That is fine for a demo. It is a liability in anything real.
A prompt is a wish
The mistake I see most often is trusting the prompt to do the enforcing. People write "only use real quotes" into the instructions and believe they have made a rule. They have made a wish. A prompt is a request to a system that is free to ignore it, and sometimes will. If a thing has to be true, you cannot ask for it. You check it after the model has spoken, in code that does not care how confident the model was.
A prompt tells the model what you want. The check is what makes it so.
None of this is glamorous. It never shows up in the demo, because the demo is the one run where the model happened to behave. It shows up on run four hundred, when the model says something plausible and wrong, and the code catches it and drops it, and the person never knows how close they came to being handed a lie about themselves.
It is the same rule I have been writing about all along, moved down a layer. Automate the work. Never automate the judgment. In a sales process, that means a person decides who is worth calling. In the code, it means the model reports and the program decides. The judgment stays where it belongs, which is never with the thing that cannot be held accountable for being wrong.
The model does the work. The judgment stays in code.
That rule runs through everything I build, from INTELLaMiNER to the voice agents behind it. Automate the work, verify the output, and never let the thing that cannot be held accountable have the final say. If that is how you want AI built, that is what I do.
intellaminer.ai