Process-based ecosystem models are used increasingly to evaluate the impacts of agricultural practices on soil organic carbon (SOC) stocks at various scales. One of the major sources of error and projection uncertainty in these models is the specification of the initial SOC pools sizes. However, few studies have examined errors and uncertainty over time and for various agricultural practices. The main purposes of our study were 1) to examine the impacts of initialization scenarios on CENTURY model V4.5 performance and 2) to quantify the initialization contribution to the total variance of error of the CENTURY model. We simulated the SOC dynamics of six well-characterized long-term experiments (LTEs) with 25 treatments across France, testing various agricultural practices (i.e., inorganic and organic fertilization, various crop rotations and straw and residues removed) using the CENTURY model while keeping the standard parameters unchanged. We applied nine initialization scenarios, each characterized by a unique combination of crop management and relaxation procedures. These relaxation procedures consisted of shifting simulated SOC and nitrogen levels at the end of the initialization period until they matched the stocks at the beginning of the experiment. At the end of the initialization period, the distribution pattern of SOC pools was similar in all scenarios for all LTEs. The slow pool represented the largest proportion of total SOC stocks (average value of 61.5%), whereas the active and passive pools averaged 5.3% and 27.9%, respectively. The overall analysis of CENTURY performance indicated fair results for SOC stocks prediction (R2 values of the nine initialization scenarios ranged between 0.50 and 0.75) but weak results for SOC change prediction (R2 values of the nine initialization scenarios, ranged between 0.1 and 0.36). The root mean square error (RMSE) values were moderate compared to the total measured SOC stocks and their confidence intervals. The RMSE v