Deterministic core, AI amplifier

On a pricing tool I built for a contractor, the AI never once computes the number that goes in front of the customer. Not ever. That was the single most important design decision in the whole system, and it's the pattern I reach for anywhere a value has to be exactly right: a deterministic core doing the work that must be correct, with the AI as an amplifier around it, never the source of truth.

The money math is plain deterministic code, the kind that does the same thing every time and can be tested to death. The model works around the edges. It reads messy inputs into clean ones, it drafts the language, it flags things that look off, it makes the whole thing faster and smoother to use. What it does not do is decide the price, because deciding the price is exactly the ninety-nine-percent problem from two lessons ago, and a wrong price isn't a rounding error, it's the one that ends the relationship. So that job lives in code that cannot roll a bad die, and the model is walled off from it by design.

How do you know what belongs in the deterministic core?

Draw the line at variance. Where variance is a feature, the fuzzy, high-judgment, many-valid-answers work like drafting, summarizing, reading intent, suggesting, the model's variability is exactly what makes it valuable, and you want it there. Where variance is a catastrophe, the exact, one-right-answer, zero-tolerance work like money, permissions, deletions, anything a customer treats as a promise, that same variability is a liability, and you wall it out. The skill is being honest about which side of that line each piece of your system is on, because the temptation is always to let the impressive tool touch the important thing.

Why "amplifier" is the right frame

Because it keeps the model doing what it's genuinely great at while removing it from what it's structurally wrong for. An amplifier makes the core louder, faster, more usable. It does not replace the core, and it can't corrupt the signal, because the signal was computed somewhere it can't reach. You get the model's speed and flexibility on the ninety percent that benefits from it, and hard determinism on the ten percent that can't survive a single miss. That's not using less AI. It's aiming it.

The takeaway: put the must-be-exact work in deterministic code and let the AI amplify around it, never compute it, because variance is a feature where answers can flex and a catastrophe where they can't.