There are currently many problems with the development and assessment of software intensive safety-critical systems. In this paper we describe the problems, and introduce a novel approach to their solution, based around goal-structuring concepts, which we believe will ameliorate some of the difficulties. We discuss the use of modified and new forms of safety assessment notations to provide evidence of safety, and the use of data derived from such notations as a means of providing quantified input into the design assessment process. We then show how the design assessment can be partially automated, and from this develop some ideas on how we might move from analytical to synthetic approaches, using safety criteria and evidence as a fitness function for comparing alternative automatically-generated designs.