🤖 AI Summary
本文研究了在不满足单交叉条件下，如何通过跳跃或集中处理具有最小有效规模技术的代理问题，并找到全局最优解。
📝 Abstract
A principal screens an agent whose technology has a minimum efficient scale, so the Spence-Mirrlees condition fails along a monotone dividing curve: the locus at which every type values marginal output equally. For the class in which this curve and the relaxed solution are both strictly monotone, the optimal contract obeys a trichotomy, governed by how the two meet: a jump is impossible when they never meet, unavoidable across a flat dividing curve, a choice across a strictly increasing one. The optimum is found, not conjectured: each solution is certified as globally optimal among all implementable allocations, deterministic or random, by dualizing the family of binding constraints through an explicit weight; the certificates require neither linear primitives nor any restriction on the shape of the contract. Under mild regularity the class comprises exactly forty configurations; each is mapped to its forced shape, solved in closed form, and certified.