Abstract
dc:description.abstractLarge language models are evaluated on whether the code they generate compiles and passes tests. Whether code is usable is a question that has received little attention so far. This thesis asks it in the context of cross-platform mHealth application migration. A sleep-tracking application (React Native, 82 files, 24 screens) was migrated to three frameworks (Kotlin Multiplatform, Flutter, .NET MAUI) by two frontier AI coding agents (Claude Sonnet 4.6, GPT-5.5) under a shared 15-rule prompt. All seven implementations were evaluated by a single assessor on an Android device using Nielsen’s ten heuristics across six standardized tasks. One of six migrations improved on the baseline while five degraded usability. Claude Sonnet 4.6 outperformed GPT-5.5 on every matched framework, but framework choice mattered more than model choice, as declarative UI translated well across all targets while imperative behavioral contracts did not. Both tools independently converged on the same catastrophic failure, suggesting this failure is driven by framework API design rather than by the model itself. The thesis contributes a severity matrix spanning two tools and three frameworks, a reusable heuristic-based evaluation methodology for AI-generated migrations, and a five-type regression taxonomy including modality substitution, a failure mode that existing benchmarks would mark as correct.
Degree
thesis:*- Department dc:contributor.department
- Computing and Software
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Mokhtari, Melvin
- Advisor dc:contributor.advisor
-
- Paige, Richard F.
Subjects
dc:subject × 8Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Identifier URI
- https://doi.org/10.71548/2228