It is equal to the Fisher Information matrix.

It is defined as $H J^{-1} H$, where $H$ is the Hessian and $J$ is the sensitivity.

It provides the sandwich estimator of variance $A^{-1} B A^{-1}$, where $A$ is the expected Hessian and $B$ is the variance of the score.

It is independent of the model choice.