The async planner is the part I'd love more detail on. If it's reasoning ahead while the kid is still talking, some chunk of that work gets invalidated every time the kid says something off-script, which at age 5 is presumably constantly. Is the speculative planning you throw away a real cost line or a rounding error? I build streaming LLM stuff (nothing this ambitious) and the "right move at the right moment" framing rings true, latency budgets change what the model can even attempt per turn.