One possible explanation is that the default convergence tolerances are too loose. You might need to tighten the tolerances (CVNTOL). With surface-to-surface radiation the radiosity solver uses a decoupled approach. The surface temperatures of the radiating surfaces are used to calculate the radiative heat flux. This calculated radiative heat flux is then applied to the FE model for the conduction analysis. With this decoupled approach, the heat flow rate vector can become ‘large’ causing the convergence tolerance to become ‘loose’. Thus, for some models, the convergence criterion (tolerance) needs to be tightened. With the default criterion, convergence can be achieved prematurely before obtaining a correct heat balance. The automatically calculated default criterion is based on the applied load, but radiation in a perfect cavity is not actually an ‘applied load’. It should cancel out (heat is not input or removed from a perfect enclosure). However, when setting the default criterion, the solver treats the radiative heat flux as an external load (since it is applied like an external load). Thus, the default criterion can be too loose to obtain convergence.