The Human Compilation Layer: Decision-Sufficient Abstraction and Traceable Verification for AI-Generated Technical Artifacts

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

As AI systems generate increasingly complex technical artifacts, the main challenge is often no longer generation but meaningful human verification. This paper introduces the Human Compilation Layer, a task-conditioned interface between machine-generated artifacts and human decision-making. The framework separates task and specification alignment, machine or empirical assurance, human-facing abstraction, and final decision use. It proposes that a useful representation should be feasible to evaluate within available human resources, preserve information necessary for the intended decision, and remain traceable to the underlying artifact and supporting evidence. The paper also identifies major failure modes, including specification mismatch, semantic omission, misleading abstraction, and automation-induced overreliance. It provides a formal framework and evaluation blueprint for future empirical studies, while making no claim that the proposed approach has yet been experimentally validated.