We can emulate pretty much any old system, but emulated "perfectly" is at best not demonstrated and at worst already proven wrong. Perfect emulation implies emulating the undocumented bugs that end up working--and there are a lot of those. Getting the correct behavior in the case of parallel components (note that this applies even to single-core systems, since dedicated hardware constructs can race with the CPU) is particularly tricky. Even without parallel issues, you can have cases where software blatantly violates the API but it works reliably and consistently, necessitating complex emulation to make it work correctly (there are some games that don't work on WINE for this reason).
Heck, even if you emulate the documented performance perfectly, you can still suffer regression behavior if the existing program just happens to rely on misbehavior of the original hardware due to less tolerant exception handling, stricter memory alignment protections, etc. in the later deployment environment.
I've seen this a lot in memory misuse / misalignment when porting between various UNIX implementations. If your development environment just happens to forgive a lot of sloppiness (cough Solaris cough), misbehavior may only show up later on other platforms.
As for the rest doing cycle exact emulation of 50 year old system is and probably always will be doable and from some PoV trivial: you can just simulate the hardware logic of the original implementation on RTL or even gate level and it still will be faster than realtime.
[Edit: "reserved for implementation" was missing from the first paragraph]
The nut here is that these systems evolve in a very haphazard way do to funding availability, scaling stresses, and changing requirements.
For the IRS, I think it would be useful to engage the US Digital Service to see if they could design a system based on a cluster of non-specific hardware with only network interfaces, that could host a 'tax computation engine' and a growable data store. While it is a 'big' problem, it is not as big as some of the problems I have seen hosted on 25,000 machines as a 'small' cluster. That is 800,000 cores (assuming 32 core machines) so for 140M tax payers that is 175 payers per machine. Something that could no doubt process every return in a single day, and run fraud analysis as well.
Clearly the government needs warehouse scale computers that their vendor (IBM) is not in a position (yet?) to deliver.