Definitely. We use a combination of source maps, manual rules, and right now, human review. Source maps let us go from the JS running on the page back to original source code that generated it. It's not perfect which is why we also have some manual rules + still do human reviews when we aren't sure if we got it right. There has been some amazing progress using GPT-3 for code generation as well (check out https://debuild.co/) which is something we are also trying to use.
I've been using here.fm for several months and it is awesome. I'd highly recommend you check here.fm out before supporting a company that rips off other people's work.
Interesting idea. Are there any performance benchmarks that you've run and can share? I'd really like to get a sense for what kind of improvements you've seen for computationally heavy tasks vs. a language like Go or Python.
I just gave it a shot to play codenames and it worked great for us. Are there restrictions on the max number of participants per game at this time? I'm thinking this could be fun for a work happy hour.