Design Exploration Across the Full Ranking Funnel

A team adds ε-exploration slots only at the final re-ranking stage (2 of the top 20 slots reserved for lower-scored items) to support long-term-value learning, and is surprised that certain content categories remain almost entirely absent from the feed months later.

1. Explain why re-rank-only exploration fails to fix this, tracing the problem back through the funnel.
2. Propose a concrete exploration design that reaches earlier funnel stages, referencing what changes at each stage.
3. What is the propensity-logging requirement this design creates, and why is it non-negotiable?