In AI development, a " seed " ensures that random processes (like weight initialization) are reproducible. If an A/B test compares two models using different seeds, the auditor cannot tell if the performance difference is due to the model changes or simply due to " random luck " in how the weights were initialized. This invalidates the test results. For a fair " apple-to-apples " comparison, the seed should remain consistent. Tuning on a training set (Option B) is standard, though it risks overfitting; however, the lack of scientific control in testing (Option C) is a more immediate risk to the integrity of the change management process.