Size an Exploration Budget for a New Bandit Arm

A video platform's recommendation re-ranker adds a new "boost slightly unfamiliar creators" action to its bandit action space, alongside the 4 existing ranking strategies. The platform serves 20 million ranking decisions per day. Historically, a brand-new action needs roughly 200,000 exposures before its estimated value is precise enough to decide confidently whether to keep it.

1. If the team wants the new action to reach that many exposures within 3 days, what fraction of daily traffic must be allocated to it, and what does that number tell a stakeholder worried about cost?
2. The new action underperforms the current best strategy by an average of 0.4 percentage points of session-continue rate during the exploration window. Estimate the daily cost in "lost session continues" at your computed exploration rate, and explain how you'd present that number alongside the value of the information gained.
3. Once the new action's posterior has narrowed and it turns out to be worse, describe the schedule you'd use to wind its exploration budget down to zero, and what should happen to that traffic instead.