A pinned local Qwen3-4B study made every compiled-context answer structurally valid while observed correctness and unsafe acceptance both rose.