Diagnosing a Position Bias Bug

After launching a new ranker, an analyst notices the model assigns unusually high scores to videos regardless of content quality whenever they appear in slot 1 of the training logs, and the online CTR curve by slot shows position 1 gets 5x the CTR of position 10 even for comparable content.

1. Explain the mechanism that causes a ranking model to learn this pattern if position is not handled correctly.
2. Describe two concrete training-time fixes and how each works at serving time.
3. How would you empirically confirm the fix removed the bias rather than just hiding it?