From question to verifiable answer

An answer is not a single operation

When an assistant answers correctly, several mechanisms have succeeded together. It understood the request, isolated useful facts, found an applicable policy, produced a formulation and satisfied the evaluation criterion. The final answer alone cannot tell us which mechanism deserves our trust.

Consider a customer who wants to return a personalized keyboard received twelve days earlier. A general page allows returns within thirty days. The assistant may quote that genuine page and still be wrong if a separate policy excludes personalized products.

Five checks, in the right order

First check intent: is the customer reporting a defect or merely changing their mind? Then check facts: personalized product, twelve-day delay and no defect. The third check finds decisive evidence, including exceptions. The fourth ensures that the answer does not claim more than the documents support. The fifth evaluates the outcome according to the real cost of an error.

This decomposition turns a global impression into a diagnosis. When decisive evidence is missing, asking for verification may be more rigorous than choosing yes or no.

A reusable method

Before writing a prompt, create a four-line brief: expected decision, required facts, authoritative source and behavior when that source is missing. Then build a nominal case, a case where one condition changes and a case where the decisive information is unavailable.

This produces a small protocol that explains why an answer succeeds or fails. It does not prove that the system works everywhere, but it provides an observable, reproducible and improvable starting point.

Remember

A convincing answer is the visible output of a chain of decisions. To build a reliable tool, test the links separately and connect every conclusion to the evidence that permits it.