We benchmarked it honestly: on nine native tasks across Swift, Kotlin, and React Native (common APIs, Metal, MediaCodec, iOS 26 and Android 16 surface), a frontier model wrote correct API calls with no tools at all. That is not the problem anymore. The problem is what confident code hides: an API deprecated at your targetSdk, a call that does not exist below your deployment target, a spelling that compiles in one context and not another. Compiled Swift and Kotlin get most of that from the compiler. Native calls made from React Native or NativeScript get none of it, and no compiler enforces a deprecation policy against the target you choose, or the one you are moving to.
sdk.run answers with verdicts, not prose: built from the same compiler-parsed metadata NativeScript uses to call these APIs at runtime, so every symbol carries its exact signature, native selector, framework, and per-version availability. It graded our own hosted oracle and found two defects our launch tests had missed. Verification does that.