I think that's the key problem: nobody has a complete big-picture idea of what the program is doing. Sure, there are people in each service branch that are familiar with subsets of the overall technology, but I suspect that at the overlap of these subsets there is significant disagreement over interpretation & implementation that has never become a visible problem because nobody runs consistency checks.
Since compiling the mutually contradictory payroll policies into a non-probabilistic master policy would immediately shed light on these issues, it would be a political minefield. Every choice you make would be the wrong choice and nobody is simultaneously high enough on the totem pole to say "live with it" and low enough on the totem pole to actually implement it.
Everytime someone tries to start a system to fix pay they realize that they will need to change all the personnel systems to match, and figure that they'll go and fix the personnel systems too.
They tried that with DIHMRS, which actually did involve billions, and the project never even got finished it was such a C.F.
Surely a team with a background in large systems could take a set of requirements and produce a working system in 12-18 months.