Bounded Research Organization Lab

This lab compares two ways to research the same large corpus: a flat Sol-to-two-scout fan-out and a depth-two Kimi-to-two-scout hierarchy. MiniMax scouts, prompts, source questions, and trial tasks remain matched so the handoff itself can be evaluated.

Kimi may spawn at most two scouts and the whole session permits only three worker launches. The root continues repository work instead of waiting. Promotion requires evidence that synthesis quality or root productivity outweighs extra latency, duplicate work, and information lost across the additional handoff.

Kimi and every MiniMax scout use persistent supervised Unix app-server threads with goal/turn lifecycle, warning-only stalls, terminal history, and retained partial evidence. The root can control either layer, while Kimi can control only its scouts; detach, pause, interruption, continuation, finalization, full stop, and cancellation preserve the bounded hierarchy instead of replacing it with polling or model-authored clock checkpoints.