Since the 1950s, citation number or “impact” has been the dominant metric by which science is quantitatively evaluated. But research contributions play distinct roles in the unfolding drama of scientific debate, agreement and advance, which are differentially valued by scientists and their institutions. Computational power, access to citation data and an array of modeling techniques have given rise to a widening portfolio of metrics that extract different signals regarding their contribution to scientific activities. Besides impact, the innovation of work and the way that it builds up the scientific discussions are taken into consideration. Here we unpacks the complex, temporally evolving relationship between citation impact alongside novelty and disruption, two emerging measures that capture the degree to which science not only influences, but transforms later work. Novelty is measured at the point of production and captures how research draws upon unusual combinations of prior work. Disruption is measured over time and captures how research comes to eclipse or amplify the prior work on which it builds. We theorize that novel papers will exhibit disruptive impact over time, and demonstrate how they are much more likely than conventional papers to disrupt current literature. Novel papers do not do so immediately, but often become “sleeping beauties”, accumulating surprising attention and citation impact over the long run. In summary, new directions for science are created from a lack of consensus. Finally, we show how novelty can be reformulated as the combination of ideas across knowledge space to reveal the combinatorial nature of advance. The evolution of knowledge space over time characterizes how yesterday’s novelty forms today’s scientific conventions, which condition the novelty—and surprise—of tomorrow’s breakthroughs.