Agonistic interactions between animals are often settled by the use of repeated signals which advertise the resource-holding potential of the sender. According to the sequential assessment game this repetition increases the accuracy with which receivers may assess the signal, but under the cumulative assessment model the repeated performances accumulate to give a signal of stamina. These models may be distinguished by the temporal pattern of signalling they predict and by the decision rules used by the contestants. Hermit crabs engage in shell fights over possession of the gastropod shells that they inhabit. During these interactions the two roles of signaller and receiver may be examined separately because they are fixed for the duration of the encounter. Attackers rap their shell against that of the defender in a series of bouts whereas defenders remain tightly withdrawn into their shells for the duration of the contest. At the end of a fight the attacker may evict the defender from its shell or decide to give up without first effecting an eviction; the decision for defenders is either to maintain a grip on its shell or to release the shell and allow itself to be evicted. We manipulated fatigue levels separately for attackers and defenders, by varying the oxygen concentration of the water that they are held in prior to fighting, and examined the effects that this has on the likelihood of each decision and on the temporal pattern of rapping. We show that the vigour of rapping and the likelihood of eviction are reduced when the attacker is subjected to low oxygen but that this treatment has no effect on rates of eviction when applied to defenders. We conclude that defenders compare the vigour of rapping with an absolute threshold rather than with a relative threshold when making their decision. The data are compatible with the cumulative assessment model and with the idea that shell rapping signals the stamina of attackers, but do not fit the predictions of the sequential assessment game.