CT-AI: Certified Tester AI Testing Version 2.0

Pre-trained models and transfer learning can introduce or preserve bias-management, reproducibility, data-quality, and security concerns, while inconsistent data preparation between source-model training and subsequent use can reduce functional performance. Given that the selected source model is already highly accurate, widely adopted, and intended for a similar use case, unexpectedly low performance of the pre-trained model is the least likely significant risk among the stated alternatives.

Machine learning models require large volumes of relevant, well-labeled historical data to learn accurate predictive patterns. An old application whose defect records are explicitly linked to failed tests and production incidents provides exactly this: a rich dataset where each defect's real-world impact (i.e., its true criticality, as evidenced by test failures and production incidents) is known. This direct traceability between defect characteristics and actual consequences gives the ML model clear, reliable labels to learn from, enabling it to accurately classify the criticality of newly reported defects. In contrast, a new application lacks sufficient historical data, a new team (despite existing defect data) does not affect the model's data-driven learning process as directly as label quality does, and a dataset skewed heavily toward non-critical defects suffers from class imbalance, which degrades the model's ability to correctly identify critical defects.

Annotation assigns labels to image content, including the position of an object such as a football through bounding boxes, keypoints, or segmentation masks.

Incorrect labels introduce erroneous ground truth into the training data, causing an object-detection model to learn incorrect classifications or bounding-box associations and thereby reducing prediction accuracy.

Data quality concerns the integrity of model input data, including issues such as missing, incorrectly typed, or out-of-range values. Hardware, sensor, and security failures can compromise those inputs. Incorrect weights are model parameters that affect how a model processes inputs and generates outputs, so they are a model-performance issue rather than an input-data-quality issue.