I lived in that world on a DoD program for 3 years. I still have nightmares.

We also had a customer who didn't want to "pay for the same code twice". That meant signal processing algorithm research had to be done in the operational real-time code, and that we could have one and only one codebase (including branches but not including engineer/scientist working copies). The "experts" from Mitre and elsewhere who were managing us were so clueless they didn't realize this was costing them more and adding more risk than allowing the scientists to do their algorithm development in Matlab and Fortran, then have the engineers turn the algorithms into an operational system. Of course, we still had to have requirements and design documentation up-front before we could implement code, so...