Multi-objective ranking with tunable weights

Design the ranking layer for a feed that must optimize multiple competing objectives at once: predicted click, predicted long-dwell, predicted 'meaningful interaction' (comment/share), and a penalty for low-quality/clickbait. Each is a separate model producing a probability; the final order must combine them into one score. Product keeps wanting to retune the balance (push 'meaningful interactions' up before an election; push dwell up to grow watch time) without a code deploy. Design the scoring and the tuning mechanism.