SHOUT-Trainer: Closed-loop Trainer for Silent Data Corruption Hunting and Observation Using Transformers

Silent data corruptions (SDCs) pose a significant reliability challenge in modern computing systems as they silently propagate incorrect results without triggering architectural exceptions or system failures. Fault injection (FI) is the de facto approach for quantifying SDC behavior, but large-scale fault injection campaigns are computationally prohibitive and difficult to apply continuously during software development. In this work, we propose a learning-based framework that predicts the probability of SDC propagation directly from source code using large language models trained on FI-annotated programs. To improve prediction accuracy, we introduce a closed-loop training strategy in which weakly predicted SDC program sites are identified and used to guide targeted FI campaigns on the corresponding programs. The resulting FI data is incorporated into the training dataset, creating a closed-loop retraining cycle. Experimental evaluation shows that the closed-loop predictor reduces the mean absolute error by 43.3% on the poorly predicted SDC program sites of the open-loop model, confirming that targeted feedback improves prediction accuracy in the most challenging regions of the program space.