This article describes a portion of a study of cognition in the wild that was conducted to investigate how human operators of nuclear power plants monitor the state of the plant to detect abnormalities. Although the study was originally motivated by applied concerns, it surprisingly led to evidence relevant to basic theories of perception. The findings reveal that the instruments in the control room that operators can use for monitoring are both fallible and limited in their informativeness, thereby occasionally providing an inaccurate indication of plant status. As a result, operators sometimes leave the control room to directly observe components in the plant using their unaided perceptual systems. Because such information is lawfully constrained and not mediated by instruments, it provides a rich, reliable, and therefore, unique and highly valued indication of the true status of the environment. These findings show that Gibson's (1979/1986) distinction between directed and mediated perception is both pragmatically and psychologically relevant. The results also have very important implications for experiments on direct perception based on computer-simulated, rather than lawfully constrained, experimental stimuli.