Can the small mind choose a useful next step?
Qwen can now send real structured decisions through the isolated simulator and save the result. Its repeated choice to visit the village shows why valid actions still need a longer behavior check.
Full entry & measurements
Real local-model compatibility probe, not Unity gameplay acceptance. Source qwen-action-probe-e92e0a210513, llama.cpp 0.6.0/build11429/d81235049, 15 October 2026 Zurich. Three of three Action-schema decisions applied and persisted; prompts excluded the other-resident private sentinel. Measured isolated ledger: 8,258 input/63 output tokens; latencies 16.40/13.10/12.68 seconds. All choices were visit:village despite varied suggestions, so purposeful autonomy remains unverified. The initial mistakenly paused fixture rejected all three attempts without mutation, preserved separately. No application source or live resident state changed; both probe workers closed. Historical approved frame reused only as bounded public illustration; no fresh camera provenance claimed. Read-only stale player-process diagnosis and these results were delivered to the active native-game chat. Fresh game-camera connection and native model-panel evidence remain pending its shared-player handoff. Details and private evidence locations in docs/local-first-refinement.md.