Models are used increasingly to predict long-term changes in soil organic matter (SOM). Comparison with measured data is clearly desirable. We compared simulations of the mechanistic ITE (Edinburgh) Forest (EF) and Hurley Pasture (HP) ecosystem models with experimental SOM data from three long-term experiments: a 30 year old pine forest in South Carolina, USA, a 100 year old area of naturally regenerating woodland at Rothamsted in southeast England, and a 140 year old grass pasture subjected to various input regimes also at Rothamsted. EF's model trees died too readily during occasional periods of drought, so we cut out the water submodel (which includes leaching): the cut-down model simulated measured accumulation of C to within around 10% but greatly overestimated that of N, when leaching was in fact significant. Again, and for the same reason (plant death during drought), we had to cut the water submodel out of HP: the resulting simulations generally overestimated SOM-N, especially in treatments receiving nitrogenous inputs, and bracketed the measured SOM-C data. Simulated SOM levels responded rapidly to organic and inorganic inputs, however, whilst measured data did not. We therefore rewrote the SOM submodel to include protected and stabilised SOM pools, in an attempt to buffer the system. The new submodel showed little effect of treatment, improved SOM-N simulations, but consistently overestimated SOM-C. This mismatch between measurement and model may reflect nothing more than tao shallow a sampling depth. We performed no site-specific parameter optimisation because: (1) the data sets are small; (2) it is not clear how much of the SOM in the system is contained within the experimental sampling depth; and (3) the models are mechanistic, with parameters reflecting real measurable properties of the system they represent. In the absence of such tuning, the models should simulate other relevant systems just as well as those presented here.