In this paper, I delineate two major problems facing reliabilist appro
aches in epistemology. I argue that Alvin Goodman's (1986) position fa
ils to solve either problem. I then suggest an alternative reliabilist
approach that ties truth-ratio assessments to particular, well-specif
ied cognitive tasks. I claim that a well-specified cognitive task is a
n empirical hypothesis about a system that involves the specification
of input and output types and nomic correlations (including statistica
l correlations) that underlie the system's performance. On my approach
, one characterizes processes by reference to the system's disposition
s across the situations consistent with the task. Characterization is
best understood as revealing a strategy or a set of strategies for gen
erating outputs from inputs relying on certain nomic correlations asso
ciated with the task.