What have you changed your mind about when it comes to what actually makes AI coding reliable?

Very tricky because everything in the AI coding space changes every two days, maybe every three days if I'm being, I don't know, nice to it.

I thought about this for a bit.

I used to think... that up-to-date documentation and a strong test suite were enough to make AI coding reliable.

But I've learned a lot from building parts of Boost and building Fuel, and as models and harnesses have evolved, and models are just too untrustworthy to rely on those.

Now I focus a lot more on automatic verification.

So I have enforced quality gates, try to be as deterministic as possible.

Tell us why this subject, why now? Go from there.