Rédiger un PREreview

Generative AI can already produce code and functionality at low cost, yet the reliable delivery of complex software

still depends on continuously managing task constraints, component dependencies, change impacts, and validation

feedback. This paper examines the complete AI development system, comprising large language models, development

agents, tools, execution environments, validation feedback, and multi-round iteration, and takes the task session—from

task acceptance to completed delivery, termination, or further decomposition—as the unit of analysis. On this basis,

the paper defines Search Space (S), Coverage Capability (C), and Delivery Reliability (R), and proposes the Search

Space–Coverage Capability–Delivery Reliability Relationship (SCR) as a foundational framework for analyzing

AI software delivery. Delivery Reliability (R) describes whether a specific AI development result can be trusted as a

completed delivery. It is neither equivalent to correctness under known criteria nor to the statistical success rate across

repeated tasks. Within the SCR framework, the match between Coverage Capability (C) and Search Space (S)

constitutes a foundational constraint on Delivery Reliability (R); when other key conditions are comparable, a more

adequate match generally makes a specific delivery more likely to reach a state that can be trusted. This yields two

basic intervention directions: expanding Coverage Capability (C), or adjusting Search Space (S). This relationship is

definitional within the framework rather than a fully operationalized, self-contained quantitative formula, and it does

not assume that Search Space (S) and Coverage Capability (C) can be reduced to single scalar values. This paper

focuses on the latter direction and defines Search Space Governance as the engineering process of continuously

adjusting task boundaries and the engineering environment to bring Search Space (S) within the range that the current

Coverage Capability (C) can consistently handle. Although Search Space (S) is difficult to measure on an absolute

scale, it is often possible to assess how local engineering changes alter the scope, structure, distinguishability among

candidate paths, and validation conditions that typical subsequent tasks must handle, and to calibrate those assessments

progressively through task trajectories and delivery outcomes. Because AI-generated code also becomes part of the

engineering environment for subsequent tasks, Search Space Governance must continue throughout project evolution.

In large, long-lived AI software projects, feature generation reflects how many direct outputs a project produces,

whereas long-term engineering governance affects whether those outputs can continue to be understood, modified, and

validated by subsequent tasks. The SCR relationship therefore provides a common structure for describing search

space, system capability, and delivery outcomes; Search Space Governance identifies what can be directly acted upon

through engineering; and the analysis of long-term projects further shows that engineering progress also depends on

whether current results can be carried forward consistently into subsequent work. Together, these elements explain

both the continuity and the change between traditional software engineering and AI software engineering, and provide

a foundational analytical framework for engineering theory, system implementation, and subsequent evidence-based

research in the AI era.