Background : A major weakness of off-line assessments of "asyntactic" comprehension is that the task is transparently related to the comprehension performance that is the researcher's interest. Not only might the patient recruit problem-solving strategies, but his/her conception of the task may have confounding effects on performance. Aim : This study investigates the nature of individual differences in two patients' performance on two off-line comprehension tasks (selected for their low task demands) by manipulating task variables. Methods & Procedures : The two single-case studies investigate, in a test-hypothesis-retest experimental design, performance on auditory meaning classification and actor identification. Both participants (NJ and CV) exhibited agrammatism following a stroke, although CV's was milder. McNemar exact (Binomial) probabilities were calculated on non-canonical versus canonical comparisons, whereas the Binomial test and '2 probabilities gauged departures from chance. Outcomes & Results: The two patients departed from baseline for different reasons: CV showed unexpected learning, and maintenance of that learning, on one task, despite washout periods; NJ's fall from baseline was only reinstated with manipulation of the instructions in between washout periods, and he failed to generalise to the second task on a subset of the items. Conclusions : It was surmised that strong induction skills in CV made for easy abstraction of "rules" from the sentence stimuli, unlike weak induction skills in NJ which led to bias and failure to detect rules despite feedback on practice items. The findings have relevance for treatment outcomes and extend Berndt and Mitchum's (1998) claim that only minimal intervention is needed for strongly analytical candidates--it may be as minimal as feedback to exemplars.