From choosing a plot to watering it
Can a resident finish a small garden task? The simulator now carries an object choice through walking and timed work, and a compact Qwen prompt has produced a real watering decision in an isolated world.
Full entry & measurements
Python daily-life increment, 15 October 2026 Zurich. One use intent reserves an object, approaches it, performs timed work and applies a completion effect. Cancellation and interrupted routes give no reward. Empty/planted/growing/ready crop stages support seed, water, growth and harvest; timers and reservations persist and pause safely. Fresh worlds alone get a shared bench and two empty planters; existing habitats are preserved. Runtime skips inference during movement/work and retains sleep consolidation. New Unity presentation is owned by the game session and remains separately pending. Measured snapshot: 344 isolated Python tests passed in 15.52s (daily-life-tests-final.xml; relevant Python source hashes in daily-life-source.json). The exact integrated compact-context Qwen3.5-2B fixture accepted 3/3 decisions: water a planted plot, visit market, visit village; watering completed. Inputs 1,449–1,470 tokens/call, latencies 7.20/3.83/3.83s with first cold, 23:08 UTC on 14 October (01:08 Zurich 15 October), installed llama.cpp 0.6.0/build 11429. This used a historical real game image and a synthetic structured world, not a fresh resident-camera session. Earlier long-context fixtures mostly repeated greetings/travel; the intermediate compact experiment omitted landmarks and completed water/rest/harvest/replant with one rejected premature use. Every result is retained privately; these are small scoped checks, not sustained autonomy or comparative benchmark claims. No new rendered visual was produced by this Python-only increment. The reused approved frame retains its original build/date and is the historical model input, not a picture of new crop stages. The live habitat stayed paused after a private backup; caps, assignments and saved lives were preserved. All isolated workers were closed, and the idle live worker was unloaded for memory headroom pending coordinated native recheck. Next: render the additive interaction/crop state and check the new runtime against a fresh game view. See docs/local-first-refinement.md.