<![CDATA[Large language models (LLMs) can assist unit test development, but faster production of candidate tests may not translate into more effective usable test suites. This study analyzes a balanced repeated-measures dataset comparing LLM-assisted and developer-written Java unit tests across mutation effectiveness, structural coverage, and development effort. The supplied records represent 32 developer profiles, 24 modules grouped into six projects, and 192 tasks, with 96 tasks per condition. The outlined study period was November 2025–February 2026 in Semarang, Indonesia; recorded task dates span December 2025–January 2026. The primary outcome assigns zero effectiveness to unusable suites while retaining strict mutation scores for usable suites. Mean effective mutation scores were 55.22% for LLM-assisted tasks and 52.54% for manual tasks. The workbook's adjusted difference was 2.68 percentage points (95% CI −2.43 to 7.79; p = .304). Usable suites were less frequent with LLM assistance (87.5% versus 94.8%), whereas conditional branch coverage was higher (81.54% versus 74.40%). Mean effort was 26.36 versus 33.00 minutes, with an adjusted geometric mean ratio of 0.787 (95% CI 0.747–0.829). These records illustrate how an apparent advantage among usable suites can diminish when delivery failures are included. Contradictory provenance labels prevent interpreting the supplied records as verified empirical observations pending source reconciliation.]]>
Copyrights © 2026
