Principled Mechanism Design with Evidence

We cast mechanism design with evidence in the framework of Myerson (1982), whereby his generalized revelation principle directly applies and yields standard notions of incentive compatible direct mechanisms. Their specific nature depends on whether the agent's (verifiable) presentation of evidence is contractually controllable, however. For deterministic implementation, we show that, in general, such control has value, and we offer two independent conditions under which this value vanishes, one on evidence (WET) and another on preferences (TIWO). Allowing for fully stochastic mechanisms, we also show how randomization generally has value and clarify to what extent this value vanishes under the common assumption of evidentiary normality (NOR). While, in general, the value of control extends to stochastic implementation, neither control nor randomization have any value if NOR holds together with WET or TIWO.