Design offline ranker evaluation

Design the logging and offline-evaluation infrastructure that lets a ranking team estimate 'how would a new ranker have performed?' without shipping it to live traffic for every candidate model. The hard problem: your logs only contain feedback (clicks/conversions) on items the current ranker chose to show, you have no labels for items it didn't show, and clicks are heavily position-biased (top results get clicked regardless of relevance). How do you log and evaluate so offline estimates predict online results?