From Generation to Governance: A Conceptual Framework for Specification-Driven Development with Large Language Models

Specification-Driven Development positions specifications, rather than code, as the authoritative artifacts from which software systems are developed. Large Language Models generate these specifications from stakeholder intent and reuse them to produce downstream artifacts and further specifications. As the specification base grows, it becomes the authoritative knowledge base guiding the development process. Yet, current workflows emphasize generation rather than governance. As specifications remain unstructured natural language, reviewers must manually reconstruct scattered domain knowledge, while language models lack explicit grounding when building on earlier specifications. Flaws can therefore be overlooked and propagated across subsequent specifications and derived artifacts. This work-in-progress proposes a conceptual framework for governing the specification lifecycle. Specifications are generated with typed domain elements embedded in their text, preserving stakeholder readability while enabling deterministic extraction. The framework derives focused views for human review and incrementally constructs a domain model that grounds subsequent generation, supports cross-feature consistency checking, and coordinates the propagation of accepted changes. Human stakeholders retain final authority over unresolved conflicts. We present the framework, a concrete instantiation, and a research agenda.