We study sequential message-sending games with an uninformed decision
maker and multiple self-interested informed agents in which the abilit
y to prove claims is limited. We give necessary and sufficient conditi
ons for the existence of robust inference rules-that is, rules which l
ead to full, correct inferences even if the decision maker has very li
ttle information about speakers' preferences or strategies. Surprising
ly little provability is needed when the decision maker only knows tha
t the speakers have conflicting preferences over his actions. Conflict
ing preferences guarantees that someone will have an incentive to ''co
rrect'' any mistaken inference. (C) 1995 Academic Press, Inc.