1. If we're using native harnesses, I'd have preferred you use kimi code, not opencode 2. The variation in the two kimi providers just shows how you can't trust n = 1 trials