The accuracy and robustness of numerical models of geologic CO2 sequestration are almost never quantified with respect to direct observations that provide a ground truth. Here, we conduct CO2 injection experiments in meter-scale, quasi-2D tanks with porous media representing stratigraphic sections of the subsurface, and combine them with numerical simulations of those experiments. We evaluate (1) the value of prior knowledge of the system, expressed in terms of ex-situ measurements of the tank sands' multiphase flow properties (local data), to obtain an accurate simulation; and (2) the predictive capability of the matched numerical models, when applied to different settings. We match three different simulation models-each with access to an increasing level of local data-to a CO2 injection experiment in tank 1 (89.7×47×1.05 cm). Matching is based on a quantitative comparison of CO2 migration at different times from timelapse image analysis. Next, we simulate a different injection scenario in tank 1, and, finally, a different injection scenario in tank 2 (2.86×1.3×0.019 m), which represents an altogether different stratigraphic section. Our models can qualitatively match the CO2 plume migration and convective mixing of the experimental truth. Quantitatively, simulations are accurate during the injection phase but their performance decreases with time. Using local data reduces the time required to history match. The predictive capability of matched models, however, is found to be similar. The sand-water-CO2(g) system is very sensitive to effective permeability and capillary pressure changes; where heterogeneous structures are present, accurate deterministic estimates of CO2 migration are unlikely.