[AnimalMuppet]

> "scope creep" is mostly just discovery failure.

Interesting. This is almost certainly right.

OTOH, there's at least two problems with the AI. One is information theory. Getting 127 requirements from two input sentences is... let's call it an unrealistically high compression ratio. There is no way that all of those requirements are actually latent in the two input sentences.

The second problem is that an AI-generated waterfall project is still a waterfall project.

What really might be useful is to treat the 127-point "requirements" as more like a due-diligence checklist. Did we think about this? Do we need that? Use it to make sure that you aren't missing things, not as the actual requirements. That is, filter it back through humans, as others have said.