[jonathanb88]

It's not primarily rules based as that would be impossible to scale. Where rules are helpful is to provide principles for when a longer or shorter phrase is chosen.

Optimising the tradeoff between more frequent repetition of vocabulary versus better quality grammar examples is where a lot of the work is.

[jonathanb88]

I'll actually expand on this, because I think it gets at an interesting point, and at a fundamental 'bet' I'm making.

The approach I use should work if there are sufficient 'scaleable' ways to improve accuracy and choose the right phrases. Language analysis, NLP, LLMs etc all help here, and there are many methods I can still use and will add over time. I'm very much at a proof of concept stage right now!

This approach won't work if accuracy improvements quickly reach diminishing returns and an explosion of rules and/or human proofing to handle edge cases are required.

I'm hoping for the former!