Quantitative Decision Policy for AI Advertising System

We have an existing AI system that manages digital advertising campaigns. We need a quantitative engineer to define a practical decision policy for when to scale, hold, reduce, or stop campaign spend. The challenge is making reliable decisions from noisy, delayed, and incomplete performance data without reacting too aggressively to short-term variance. Scope Using a representative campaign scenario and sample data we provide, define: - key inputs and metrics - evaluation windows and minimum-data requirements - confidence thresholds - scale / hold / reduce / stop logic - handling of delayed conversions - safeguards and rollback conditions - basic exploration vs. exploitation logic - shadow-mode validation before automation Bayesian methods, sequential testing, contextual bandits, or another suitable quantitative approach may be used. Deliverable A concise decision-policy specification our engineering team can implement, including: - recommended methodology - decision rules and thresholds - one worked example using the sample data - validation and safety considerations - Python pseudocode where useful Purpose We want to validate the quantitative approach before investing engineering time in the production implementation. This is a small fixed-price task, not a full advertising optimization system or production integration. Please briefly describe one relevant system you have built involving automated decisions under noisy or delayed feedback.