AI can write code. The second problem is that it forgets. Every session starts with no memory of what the system was supposed to be. Without a persistent, machine-readable description of intent, output drifts from session to session — structurally incoherent code produced at generation speed, across boundaries no review process can match. The failure mode is not incorrect code. It is correct code that no longer serves the system it was meant to be part of. This paper presents Generative Specification (GS): a programming discipline built around one structural inversion. The specification is the primary artifact. Code is derived from it, by a stateless reader, every time. A complete GS specification describes what the system is, what it must do, what it must never do, and how every behavioral claim is verified — with enough precision that a reader carrying no prior context can derive any correct implementation state. The session boundary is no longer a loss event. GS is characterized as the pragmatic programming tier: the discipline defined by derivability — what a stateless reader can correctly determine from the artifacts alone. Seven measurable properties operationalize this standard. The paper documents the methodology across six production deployments spanning brownfield takeovers, greenfield builds, and a regulated multi-tier data platform; a controlled multi-agent adversarial study establishing derivation quality as a direct function of specification completeness; and a practitioner study with 59 participants confirming the discipline transfers beyond its author. The obligation cascade the methodology produces — you do not write code, you do not read it, you do not manage infrastructure, you do not diagnose bugs — is demonstrated in the production record. The tier at which the system maintains itself is operational. The practitioner who can state what a system must be is not being replaced. They are being elevated to the problem that was always harder.
Juan Carlos Ghiringhelli (Tue,) studied this question.