This is wonderfully delightful. I wish I'd thought of it.
I am exploring decompilation - MAME ROM to idiomatic JavaScript. It's going pretty well - the LLM does a good job translating from one language (machine) to another (JavaScript) which I guess is how the LLMs got their start.
Does your input have symbols to assist understanding, or do the LLMs have to work it out for themselves?
Warning: self-promotion: https://github.com/qarl/arcade-js
Thanks! I use sci-tools for the decompilation. It produces a LISP-ish S-expression source. I also extended it to produce a JSON AST.
AH. And I see that sci-tools has access to English derived symbols from the classes and methods.
That answers my question - whether your system is generating novel symbols.