The 14C contents and the carbon (C) pool sizes for Rothamsted Carbon (RothC) model pools and soil fractions determined by physical separation and subsequent chemical oxidation were directly compared for an agricultural topsoil (Dystric Cambisol) from a long-term experimental field in central Japan. The RothC model was run from 1975 to 2009 and the C concentration and Δ14C value of each pool and for the bulk soil were compared with the physical fractions that have been, according to literature, suggested as corresponding with model pools. When compared with the default estimate of the inert organic matter (IOM) pool size derived from an exponential dependency on total organic carbon (TOC), direct estimation of IOM pool size determined from the C remaining after chemical oxidation of the silt plus clay (< 63 µm) fraction improved the model performance for TOC changes over the 34-year experimental period. The size of each model pool matched that of the physical fractions well, whereas the Δ14C values clearly differed between modelled pools and corresponding physical fractions. The physical fractions showed more negative Δ14C values with 50–190‰ (indicating slower turnover rates) compared with the active pools in RothC, whereas using the oxidation residue generated more positive Δ14C values with 700‰ (faster turnover) than the original RothC IOM pool. These results suggest that (i) the proposed physical fractionation scheme improves the model performance in simulating TOC changes at the temporal scale of decades under the current climate and (ii) long-term predictions should be interpreted carefully because of the large discrepancy in 14C-based turnover rate between the physical fractions and RothC model pools.