Llm

What 204 questions asked to both Jev and Claude showed: Jev is almost never confidently wrong, it is unsure on the questions that need careful reading, and the data says almost nothing about how well it catches problems.

How the local Bonsai model did as the executor over 28 cards: the logic was almost always right, and almost every failure was code that did not compile. And the mistake I made in how I ran it.

How Claude, Jev and a local Bonsai model are put together into one pipeline, and what its first run showed: most failures that looked like the executor’s were mine.