The model must never compute the number

Module six draws the hardest line in the operating system, and it is the one that protects you from the model's nature. There is work where the answer must be exact, the money math, the security check, the invariant that cannot flex. That work goes in deterministic code, and the AI never computes it. The model amplifies around the exact core, it does not produce the number. Cross this line and you have built a calculator that hallucinates, which is fine until the one time it matters, and the one time it matters is the only time anyone remembers.

The reasoning comes straight from how the model works. A language model produces the most probable continuation, and probability means variance. Variance is a feature where answers are allowed to flex, a phrasing, a suggestion, a draft. It is a catastrophe where they are not, a payment total, an access decision, a legal threshold. So you split the system by that property. The parts that can flex go to the model, which is brilliant at them. The parts that must be exact go to code, which is the only thing that can guarantee exact. Your estimator is the clean example: the AI helps around the edges, and the money is computed by deterministic code that the model is structurally not allowed to touch.

Why this beats just asking the model to be accurate

Because accuracy you request is a probability, and accuracy you enforce is a guarantee. Ninety-nine percent right sounds like a margin and is actually a countdown, because probability compounds across runs and the failure you did not prevent eventually arrives. For anything where a single wrong answer is unacceptable, you do not lower the odds of a wrong number, you remove the possibility by never letting the probabilistic thing compute it. The model is an amplifier bolted to a deterministic core, not the core itself.

The Field Manual's guardrails track drills deterministic-core and making bad states impossible. Here, it is the line that keeps the whole system trustworthy where trust is not optional.

The takeaway: Where the answer must be exact, put it in deterministic code and let AI amplify around it, never compute it. Variance is a feature where answers can flex and a catastrophe where they can't, so remove the possibility instead of lowering the odds.