A VERIFICATION-CENTRED FRAMEWORK FOR GENERATIVE AI-ASSISTED CODE GENERATION, TESTING, AND DEVELOPER PRODUCTIVITY

Abstract

Generative artificial intelligence (GenAI), particularly large language models (LLMs), is changing software engineering by assisting with code completion, function synthesis, test generation, debugging, documentation, and code transformation. These capabilities can reduce routine effort, but model-produced code may contain functional defects, insecure patterns, outdated API usage, or plausible explanations unsupported by execution evidence. Productivity gains therefore depend on more than generation speed: they must be assessed alongside correctness, test adequacy, maintainability, security, and developer understanding. This paper proposes a verification-centred framework for integrating GenAI into software engineering workflows. The framework combines task and repository context, model-assisted generation, static analysis, automated tests, security checks, human review, and feedback telemetry. It defines evaluation dimensions and a reproducible protocol for comparing conventional and AI-assisted workflows without treating acceptance or subjective confidence as substitutes for correctness. The paper synthesizes established research on code generation, benchmark evaluation, automated program repair, test generation, and developer studies. It also specifies metrics for functional pass rate, test coverage, review effort, change size, defect density, security findings, and task completion time. Since no model API, repository, or experiment log was supplied with the title, this manuscript does not invent measured productivity improvements or claim to have run a comparative LLM experiment. Instead, it provides a publication-oriented conceptual framework, a system architecture, equations, and an experimental design that can be implemented with public benchmarks such as HumanEval or repository-level tasks. The central argument is that GenAI produces dependable engineering value when generation is coupled to executable verification, transparent limitations, and accountable human oversight.