Keeping Engineers in the Loop: Explainable Requirement Defect Screening Across Software Requirements Specifications

Abstract

Requirements written in natural language frequently contain ambiguity, incompleteness, and statements that are difficult to validate consistently. This experience report presents a human-centered requirement-validation workflow built around a heterogeneous source corpus of 79 software requirements specification documents. The workflow extracted 9,165 requirement-like statements from this corpus and curated a manually reviewed, single-reviewer adjudicated gold set of 110 requirements, which is the basis for all reported evaluation results. Three screening conditions were compared on this gold set: the original seed heuristic, a flat rule-based screen, and an explainable precedence-based screen that returns a defect label, rationale, source span, rewrite guidance, and confidence. The seed heuristic achieved 0.573 accuracy and 0.499 macro-F1, the flat rule screen achieved 0.500 accuracy and 0.478 macro-F1, and the explainable screen achieved 0.855 accuracy and 0.891 macro-F1. The main finding is that explanation-oriented screening substantially improves practical review support in this corpus, while the remaining errors are dominated by formatting artifacts, truncated list boundaries, and contextual judgments that still require human reviewers. The contribution is not autonomous requirement assessment; it is an auditable human-in-the-loop workflow for prioritizing requirement-quality review.