Why can a perfectly accurate headline number still support the wrong decision? Because a summary can discard the very differences that matter for the answer. This anonymous, unrefereed preprint develops an aggregation licence for detecting that problem before a compressed measure is used beyond its scope. The core test is simple: if two admissible worlds share the same reported aggregate but require different target values or actions, the aggregate alone cannot determine the target. The paper separates four questions that are often blurred together: local point identification, exact global factorisation, best worst-case prediction, and decision regret. It then asks whether the required rule is measurable, stable, computable, learnable, and valid for the declared model class. Economics, fisheries, and epidemiology provide distinct illustrations. A stable catch-per-unit-effort index, for example, can coexist with very different fish abundance when catchability changes. A sealed synthetic workflow shows CPUE alone failing, an independent survey bringing ambiguity within the registered tolerances, and a later catchability shock revoking that licence. The public package includes the paper, three technical dossiers, source and claim ledgers, executable collision checks, positive and negative controls, a frozen protocol, tests, and replay receipts. These are producer-side checks. They are not field validation, independent reproduction, specialist review, peer review, or evidence that the method improves real decisions. This synthetic-voice briefing is a communication aid, not additional scientific evidence.