I'd expect that adding a rule with arch=64 on a 32bit machine
would fail.
But, arch=32/64 doesn't look like the right solution. We are exposing
the underlying architecture which is more granular that 32 vs. 64 bit.
It includes various architectures as well. Why not keep this value
the same as the output in the audit message? And if it's done as it
currently is, the records could (theoretically) be parsed on a machine
with a different cpu arch than the machine that generated the record.
Can you post a couple of examples of what the auditctl rules would look
like?