Artificial intelligence (AI) is becoming integral to manufacturing, healthcare, and autonomous systems, creating an urgent need for reliable deployment across diverse computing platforms. Yet dependable deployment remains difficult because the software systems that adapt learned models to target devices are complex and fragile. An AI model that appears valid at a high level can still fail during deployment because of hidden resource limits, data layout requirements, and platform-specific transformations. These failures are especially concerning because they may silently alter outputs rather than cause visible crashes, making them difficult to detect, diagnose, and prevent. The project's novelties are new testing foundations that uncover hidden sources of deployment failure and assess whether AI systems produce consistent results across diverse hardware platforms. The project's broader significance and importance are improved reliability and trustworthiness of AI systems deployed in high-impact settings. The project also creates interdisciplinary educational opportunities through open tools, curriculum materials, and training that strengthen workforce development in dependable AI systems and heterogeneous computing. This project develops a cross-layer framework for testing the software stack that translates and executes deep learning models on heterogeneous hardware. The research has three integrated components. First, it develops methods for automatically mining parameterized constraints from model specifications, system implementations, and hardware resource limits, thereby exposing implicit assumptions that existing testing techniques do not capture. Second, it introduces targeted constraint negation as a new form of test guidance, driving test generation toward failure-prone regions while filtering out invalid inputs before expensive end-to-end runs. Third, it develops equivalent model rewriting and backend-aware differential checking to detect latent inconsis