1: \begin{abstract} The goal of this paper is to supplement

2: the large deviation principle of the Freidlin--Wentzell theory on exit problems for diffusion processes 

3: with results of classical central limit theorem kind. We describe a class of situations where

4: conditioning on exit through unlikely locations leads to a Gaussian scaling limit for the exit distribution. Our results are based on Doob's $h$-transform

5: and new asymptotic convergence gradient estimates for elliptic nonlinear equations that allow to reduce the problem to the Levinson case.

6:  We devote a separate section to a rigorous and general discussion of $h$-transform.

7:  \end{abstract}