"Attribute-framing bias" reflects people's tendency to evaluate objects framed positively more favorably than the same objects framed negatively. Although biased by the framing valence, evaluations are nevertheless calibrated to the magnitude of the target attribute. In three experiments that manipulated magnitudes in different ways, we examined to what extent encouraging speeded or accurate responses affected the bias of evaluations and their calibration. Results revealed a dissociation between the biasing effect of framing valence and the calibrated effect of magnitude. The bias was increased in the speeded conditions relative to the accurate conditions. However, the calibration was affected by the speed-accuracy manipulation only in negative and not in positive framing conditions. We discuss the advantage of "fuzzy-trace theory" in explaining these results, suggesting that gist representations elicit the bias, whereas verbatim representations allow calibration. However, the relative contributions of these representations to evaluation vary with task demands such as speed-accuracy requirements.