In fact, the best thing to compare it to, in my mind, would be RarVM bytecode: another ISA designed to be "forward-compatible" in the sense of ensuring that even the first interpreter will be able to run code from years in the future. Nobody ever sees RarVM bytecode; nobody even realized it existed for decades. It was just there, an implementation detail of the RAR encoding/decoding process, doing its job of enabling the RAR format to change its algorithm over time.
And at least loop-unrolled x64 code describes what it's doing straightforwardly, instead of throwing around redundant mathematical abstractions and hoping the interpreter will magically pattern-match them away.
As an aside, though:
> Abstract ISAs are still languages.
Not always true. A lot of current abstract ISAs are (because they were designed to be programmed in, as well as as compile targets), but many aren't. Most code read into a modern VM gets chewed on a bit more from its "canonical" form; the internal representation that results, with threaded code and JIT profiling hooks and tracing et al, basically looks like Nock: a graph of numbers. For most VMs, those numbers are just pointers to structs it has allocated, so they can't really "live" outside the VM. Nock just goes a bit further and says "but if you give those struct-pointers persistent wire-representable identifiers ala CapnProto, you can serialize the whole internal-VM-state graph in a portable way—and then that can serve as the VM's ISA."
And no matter the form you turn a language into, it's still a language. Speech and text are generally considered equivalent in that sense, for example- it's the abstract form that matters here.
A Smalltalk live-image, a core-dump of a process, an SQLite database, etc. are all "languages" by definition 1, but obviously not "languages" by definition 2. Nock's ISA is more like those: a format for machines to generate and other machines to consume, and for humans to use tools to introspect; not a format targeted at direct human production or consumption, even through a 1:1 mapping ala disassembly.
Doesn't this ignore the concept of Urbit's "jets"? You write code that does things in a stupid-but-canonical way; it's then the runtime's responsibility to take canonical patterns and provide optimized implementations.
The cool thing is that you get to write a naive interpreter for all existing code in a couple-hundred lines, while also being able to write a good interpreter that makes the same code run quickly. Like regular Lua vs. LuaJIT, but much moreso.