A Unique-Cause-First Masking Extension for Minimal MC/DC Truth-Vector Suite Generation With Real-Code Evaluation

초록

Modified Condition/Decision Coverage (MC/DC) demonstrates that each atomic condition in a compound decision can independently affect the decision outcome. The Unique-Cause interpretation is highly auditable because only the target condition changes in each witness pair, but it can be too restrictive for source-code decisions with mixed logical structure, complement predicates, repeated occurrences, or feasibility constraints. This paper studies Boolean-expression-level MC/DC truth-vector suite generation and proposes a Unique-Cause-first Masking extension. The method first constructs a Unique-Cause seed suite, derives conservative Masking candidates using partial-space invariance, and uses a vector-count-first weighted selection model: full coverage is enforced as a hard constraint, selected-vector count receives strict priority, and a secondary penalty favors Unique-Cause witnesses and smaller masking sets. On 14,422 extracted compound decisions and 12,426 unique project-expression pairs from the Python standard library and seven open-source packages, pure Unique-Cause generation achieved full MC/DC coverage for 12,410 expressions, whereas the proposed method achieved full MC/DC coverage for 12,425 expressions. It recovered 15 expressions that pure Unique-Cause did not cover and reduced selected-vector counts for 95 additional suites while maintaining a mean Unique-Cause retention ratio of 0.9916. Controlled size-sensitivity results for decisions with 5, 10, 20, and 30 atomic conditions, Boolean-level mutation analysis, and state-of-the-art (SOTA)-inspired satisfiability (SAT)- and binary decision diagram (BDD)-based baselines further show that the method matches the SAT-style exact baseline in selected-vector count over the same conservative Masking candidate set while retaining more Unique-Cause evidence. The evaluation is limited to Boolean-expression witness selection; concrete input synthesis and source-level mutation execution remain future work.

키워드