Could we perhaps be a bit gentler here and instead of fully discouraging custom experiments, just say that they carry some risk? I don't want to entirely discourage someone from trying something in their own environment that could potentially be contributed back to the client, but agree we need to say that it carries risk. Thoughts?

I don't feel good with that idea - by themselves, experiments don't allow you to do anything you couldn't already do in userland without them. They just encapsulate and hide that dangerous stuff happens.

So if someone were to create something like this in userland, I think I'd prefer if they would do

const client = new ApolloClient()
client['queryManager'].prototype.newMethod = () => { /* ... */ }

which makes it very visible that this is a very flimsy and potentially dangerous thing to do.
I don't really want experiments to live outside of our control, since when we change any of our internals (e.g. adding newMethod on our side), it might break experimentes, or get broken by an experiment. We can only guarantee that that doesn't happen if we are the only people writing and distributing them.

We talked this through on video :)