We present an agent-based model of paper publication and consumption that allows to study the effect of two different evaluation mechanisms, peer review and reputation, on the quality of the manuscripts accessed by a scientific community. The model was empirically calibrated on two data sets, mono- and multi-disciplinary. Our results point out that disciplinary settings differ in the rapidity with which they deal with extreme events—papers that have an extremely high quality, that we call outliers. In the mono-disciplinary case, reputation is better than traditional peer review to optimize the quality of papers read by researchers. In the multi-disciplinary case, if the quality landscape is relatively flat, a reputation system also performs better. In the presence of outliers, peer review is more effective. Our simulation suggests that a reputation system could perform better than peer review as a scientific information filter for quality except when research is multi-disciplinary and in a field where outliers exist. This work was partially supported by the COST Action TD1306 ”New frontiers of peer review” (www.peere.org), by the FuturICT 2.0 (www.futurict2.eu) project funded by the FLAG-ERA JCT 2017, by the Spanish Ministry of Science and Innovation Project TIN2015-66972-C5-5-R and by the University of Valencia under grant UV-INV_EPDI17-548224. Peer reviewed