Interesting benchmarks. One thing I'd love to see separated out: how much of the gap is syntax vs. how much is the model just having seen far less of the "fixed" form in training. A simpler grammar with zero corpus behind it could easily perform worse in practice.

Congrats on the launch. The thing I keep wondering with "company brain" products is the staleness problem — context that's right on Monday is quietly wrong by Friday, and nobody notices until a decision gets made on it.

The missing primitive in most setups I've seen: per-task scope contracts. A CLAUDE.md covers what the codebase is, but a per-task declaration of what must not change and where the agent stops is a separate thing. I've started putting those as a header in the task description itself, separate from persistent context.