Abstract: It was formerly argued that alternative evolutionarily stable strategies (ESSs) are possible for animal contests characterized by some asymmetry that can be perceived with perfect accuracy. Where roles A and B refer to the asymmetry between opponents, ESSs are: [`]fight when A, retreat when B', and vice versa. Either can be an ESS, but only if the [`]reserve strategy' (=what an animal does when it fights) is sufficiently damaging. We examine the [`]war of attrition' (winner = opponent that persists longer). In a population at either ESS, reserve strategy is never normally shown; it is therefore subject to drift unless the selective action of rare individuals which break the convention is considered. These could arise either by mutation or by mistakes in role assessment. When mutations and mistakes simply specify that occasionally an animal fights when it [`]should' retreat, selection adjusts reserve strategy to a level where only one ESS (the [`]commonsense' ESS) is possible, if the asymmetry is relevant to payoff. Thus for asymmetries in fighting ability or resource value, the individual with the lower score will retreat. However, we are particularly concerned with cases where both payoff-relevant aspects (fighting ability and resource value) are asymmetric. If opponents sustain contest costs at rates KA and KB, and their resource values are VA and VB, an [`]optimal assessor' strategy defined by the interaction between the two asymmetries, is a unique ESS. It obeys the rule [`]fight on estimating role A, where VA/KA>VB/KB; retreat in B'. If mistakes can occur in both roles, but are very rate, the ESS is not fundamentally altered though there will be infinitesimal tendencies for persisting in role B. Selection to improve assessment abilities intensifies as abilities improve, but is weak if roles A and B are rather similar. Over a range of similarity between roles, an [`]owner wins' convention may be adopted if ownership correlates positively with role A and an individual cannot tell when it would otherwise pay him to break the convention. We also examine a contest in which information about roles can be acquired only during a contest itself, and at a cost. Much depends on the rate at which information is acquired relative to the rate at which costs are expended, and on whether contests normally escalate in intensity, remain at the same level, or de-escalate. Selection favours short contests when costs are high relative to resource value, where the outcome of a round contains much information about fighting ability, and where the actual disparity in fighting ability is large.