Bolwin4 hours ago
Two things
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