On‑demand delivery platforms compensate riders via incentive activities whose tiers are derived from recent completions of riders with similar histories. Operators often need ad‑hoc plans for holidays or bad weather, adjusting rider populations, payment rules and budgets, while randomized trials are scarce and time‑consuming. We introduce a request‑driven framework that composes four stages—conditional prediction, population reduction, trajectory integration and budget allocation—through seven interchangeable modules. These modules exchange conditional trajectory laws, yielding award probabilities and award‑marked moments that support any activity rule. A response‑correction step reweights abundant no‑offer histories to match the moments of a short pilot. We prove that, given a fixed menu of plans and known stage errors, the end‑to‑end value loss is bounded by the sum of the four stage errors, and each stage can be a necessary component whose omission creates an irreducible error floor. Experiments on 3,000 riders across 45 weekly origins answered all 127 weekly windows 11.04× faster with at most 0.92% value loss. On 24 new controlled response laws, a one‑week pilot with response correction reduced regret by 51.2%, while a four‑week pilot with exact summation achieved a regret gap of only +0.007 compared to an 18‑week trial. In registered studies where windows, populations, rules and binding budgets vary per request, the framework’s regret stays below that of an equally sized randomized trial and below dose‑interpolation of the same pilot data. Reusing its one‑off preparation answered 60 requests 14.1× and 2.70× faster with identical answers. Against a nine‑offer trial fitted with the framework’s own dose curve, one‑week regret is lower by 0.055.
Review