Multi-dimensional readiness signals for structured and LLM-assisted requirements inspection

Abstract

Effective software quality assurance requires that development teams inspect large collections of requirements before implementation begins, yet no existing tool helps them decide where to focus their limited inspection effort. The inspection risk of any given requirement is not a property of that requirement in isolation. Rather, it is determined by other factors, including how each requirement relates to others in the requirements specification document, how structurally central it is, how clearly it is written, and how likely it is to change. Such factors are invisible when requirements are examined individually. Moreover, the prevailing practice of collapsing these dimensions into a single weighted score through compensatory aggregation destroys the diagnostic value of each dimension independently. This thesis proposes and empirically evaluates a two-layer inspection support system. The first layer, Prioritised Requirements Inspection through Signal Metrics (PRISM), computes three collection-level readiness metrics (structural centrality, linguistic specificity, and volatility exposure), and combines them through Pareto dominance-based structuring. Rather than collapsing these orthogonal metrics into a single scalar score, PRISM preserves each dimension independently, yielding a multi-dimensional inspection ordering. Evaluated on 262 requirements, the metrics are statistically independent, and the ordering aligns significantly with independent human inspection judgment (Spearman ρ = −0.478, ρ = 0.0037), outperforming all single-metric baselines. The second layer, PRISM-Copilot, instantiates the Signal-Conditioned Inspection Prompting (SCIP) architectural pattern, which leverages the validated PRISM metrics to condition an LLM-based inspection assistant by directing each requirement toward the type of scrutiny most relevant to its specific risk profile. Rather than subjecting all requirements to undifferentiated general-purpose inspection prompts, SCIP maps each requirement’s signal profile to a targeted inspection policy, focusing the LLM’s reasoning on the concern dimensions where collection-level evidence indicates the greatest risk. A proof-of-concept study on 50 requirements demonstrates that this targeted signal conditioning causes the LLM to raise change-sensitivity concerns where volatility metrics support them and suppress them where they do not, providing empirical evidence that collection-level metrics can meaningfully steer LLM inspection reasoning. The two layers constitute a principled, end-to-end inspection support architecture that is context-aware at the collection level rather than the requirement level, establishing a practical foundation for both systematic inspection prioritization and signal-conditioned AI-assisted inspection reasoning in large-scale software engineering practice.