This is interesting — especially the focus on reducing incorrect outputs rather than just improving generation.
Feels like a common pattern across AI tools right now:
getting something generated is easy, getting something reliable enough to act on is the hard part.
Have you noticed whether developers trust the output more when the source of context is explicit (like SDKs), vs inferred?