# validity of instrumental variables

How does one check the validity of instruments for a just-identified model, since the Sargan test can only be performed if the model is overidentified?

The situation is in some sense similar to OLS - OLS is such that residuals and regressors are by construction orthogonal to each other (i.e., $X'\hat{u}=0$), so you cannot use that lack of correlation to argue that the assumption required for causal interpretation of OLS coefficients, namely $E(X'u)=0$, does hold.