Access-Efficient Escalation Lab

This lab tests whether cheap or local work can reduce scarce-tier use without unacceptable quality loss. Luna remains the responsible root; local Qwen receives only runtime-generated literal operations; DeepSeek Flash handles one objective isolated patch; GLM is a deliberate read-only escalation.

The local route is optional. Weak-model tasks have no free-form prompt surface, no network, no writes, no spawning, a single GPU slot, and hash/line verification. The lab is not promoted by lower nominal cost alone: matched trials must show at least 50% lower API-equivalent or scarce-tier consumption with no more than two points of success loss.

All three workers use persistent supervised Unix app-server threads. The root can inspect, trace, steer, interrupt, pause, continue, detach, finalize, fully stop, or cancel each one. Goal token ceilings or clock-free turns replace per-spawn model timeouts; stall warnings never interrupt, and suspension retains terminal history, trace, partial evidence, and any isolated patch. The Qwen role remains low-trust and cannot control or spawn anything.