Randomized trial examines the effectiveness of Generative Specification in improving software architecture clarity, indicating its potential to mitigate architectural drift.
The dominant failure mode of AI-assisted software development is architectural drift: an agent that cangenerate a system in one session resolves a thousand implicit decisions silently, and they diverge acrosssession, team, and service boundaries. The cause is structural — the AI is a stateless reader with nomemory, shared context, or ability to ask. Generative Specification (GS) removes the freedom to leavearchitectural intent implicit, requiring a specification from which a stateless reader can derive correctoutput, operationalized as seven scored properties (Self-describing, Bounded, Verifiable, Defended,Auditable, Composable, Executable). A layered empirical program establishes distinct, individuallycheckable claims (AX, BX, EX, KX, RX; ALX at the formal, machine-checkable tier; two single-shotexperiments, MX and RND-1), with all evidence public at github.com/jghiringhelli/generative-specification(ALX at github.com/jghiringhelli/loom). This record bundles the white paper (primary) with its Field Guide, Experiment Supplement, scoring rubric,and the full ~115-page Compendium. This version supersedes 10.5281/zenodo.19637142
No takes yet. Share an insight, caveat, or question.
Juan Ghiringhelli (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: