Output difference between linear and nonlinear model

I noticed, that the Dynare output of a nonlinear model and it’s log-linearized version don’t exactly coincide. Meaning, that considering the log-linearized value of a variable from the nonlinear model in the form of: \hat{y} = log(\frac{y}{yss}), the value is close, but not exactly the value from the log-linearized version of the model.

In the perfect_foresight context, what does Dynare do differently, compared to doing the log-linearizations by hand and running the linear model?

test_linear.mod (506 Bytes)
test_nonlinear.mod (985 Bytes)

When using perfect_foresight_solver, Dynare will not do any approximation but rather will solve the full nonlinear system. Of course the results will differ from the linearized version.