一句话：提出后指标时代用仓库挖掘、LLM 语义解释与结构度量三角互证来度量软件质量，并将信号分歧形式化为类型化诊断信号。
ISO/IEC 25010 定义了八个产品质量特性，但数十年来质量定义与实际度量之间始终存在鸿沟。本文认为根源在于以往主要依赖结构指标，难以捕捉语义质量维度。为此，作者提出 AI 辅助三角互证愿景：融合仓库挖掘的演化信号、LLM 智能体的语义解释以及经验证的结构度量。核心是将信号不一致形式化为经校准、置信度感知的分歧，并转化为类型化诊断信号，指明需关注的质量维度及应采取的行动。方法上还设计了结合规则与受限 LLM 判断及弃权的 ISO 映射机制，以及与验证和修复工作流集成的架构。若实现，质量评估将从晦涩数字转变为可操作、不确定性感知的指导。
ISO/IEC 25010 defines eight product quality characteristics, yet after decades of research, a persistent gap remains between quality definitions and practical measurement. We argue this gap exists not because the problem is intractable, but because previous approaches relied primarily on structural metrics, which struggle to capture semantic dimensions of quality. This paper articulates a vision for closing this gap through triangulation: combining evolutionary signals from repository mining, semantic interpretation from LLM agents, and structural analysis from validated metrics. Central to this vision is a formalized concept: when signals disagree, the disagreement becomes a typed diagnostic signal, revealing not just uncertainty, but which quality dimension requires attention and what action is appropriate. We present an initial formalization of disagreement as calibrated, confidence-aware divergence across signals, an ISO mapping mechanism that combines rules with constrained LLM judgments and abstention, and a triangulation architecture that integrates with verification and repair workflows. Triangulation serves as a Verification-Repair substrate: an evidence layer that gates verification tasks and guides repair actions. If this vision succeeds, quality assessment transforms from opaque numbers to actionable, uncertainty-aware guidance.
@inproceedings{AlmonteN26,
title = {Rethinking Software Quality Measurement: A Vision for AI-Assisted Triangulation in the Post-Metrics Era},
author = {Jomar Thomas Almonte and Nathalia Moraes do Nascimento},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3805580}
}